Add Batch c3bc21d9-4689-4ab6-8491-9c3f39e78182 data

.gitattributes

@@ -5155,3 +5155,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/AutoMoE_[[:space:]]Heterogeneous[[:space:]]Mixture-of-Experts[[:space:]]with[[:space:]]Adaptive[[:space:]]Computation[[:space:]]for[[:space:]]Efficient[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/0f885038-127c-477b-b9c4-7470ada93495_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Automated[[:space:]]Refugee[[:space:]]Case[[:space:]]Analysis_[[:space:]]An[[:space:]]NLP[[:space:]]Pipeline[[:space:]]for[[:space:]]Supporting[[:space:]]Legal[[:space:]]Practitioners/4db6ff2d-b617-4895-898e-bf41b1329760_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Automatic[[:space:]]Identification[[:space:]]of[[:space:]]Code-Switching[[:space:]]Functions[[:space:]]in[[:space:]]Speech[[:space:]]Transcripts/1d6ee594-ef2e-4856-b39c-4ec939a75e68_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Automatic[[:space:]]Named[[:space:]]Entity[[:space:]]Obfuscation[[:space:]]in[[:space:]]Speech/f423cca1-1aeb-43bd-8d25-5845a22e4c4c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Automatic[[:space:]]Readability[[:space:]]Assessment[[:space:]]for[[:space:]]Closely[[:space:]]Related[[:space:]]Languages/44092f05-688c-428e-a5e8-3637a992286e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Automatic[[:space:]]Table[[:space:]]Union[[:space:]]Search[[:space:]]with[[:space:]]Tabular[[:space:]]Representation[[:space:]]Learning/7244320a-8cb5-4ef7-87f0-c4d2dc4ad4ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/AxomiyaBERTa_[[:space:]]A[[:space:]]Phonologically-aware[[:space:]]Transformer[[:space:]]Model[[:space:]]for[[:space:]]Assamese/86b88357-8c3a-4270-bb19-cd2104e1901b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/B2T[[:space:]]Connection_[[:space:]]Serving[[:space:]]Stability[[:space:]]and[[:space:]]Performance[[:space:]]in[[:space:]]Deep[[:space:]]Transformers/304a7eb6-4126-4acb-9ec0-8da7fd06045b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/BLOCSUM_[[:space:]]Block[[:space:]]Scope-based[[:space:]]Source[[:space:]]Code[[:space:]]Summarization[[:space:]]via[[:space:]]Shared[[:space:]]Block[[:space:]]Representation/bdc012ed-0e12-4395-93e5-cef4ae1613b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Balancing[[:space:]]the[[:space:]]Effect[[:space:]]of[[:space:]]Training[[:space:]]Dataset[[:space:]]Distribution[[:space:]]of[[:space:]]Multiple[[:space:]]Styles[[:space:]]for[[:space:]]Multi-Style[[:space:]]Text[[:space:]]Transfer/e52f8ce0-4325-4d7a-9b9b-68ca012cb308_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/BanglaBook_[[:space:]]A[[:space:]]Large-scale[[:space:]]Bangla[[:space:]]Dataset[[:space:]]for[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]from[[:space:]]Book[[:space:]]Reviews/28378eaf-0a48-458e-a698-291e5038e7bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Benchmarking[[:space:]]Diverse-Modal[[:space:]]Entity[[:space:]]Linking[[:space:]]with[[:space:]]Generative[[:space:]]Models/6aa5cd7b-2bb5-456d-89a4-5cab0ec3e4fb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/BertNet_[[:space:]]Harvesting[[:space:]]Knowledge[[:space:]]Graphs[[:space:]]with[[:space:]]Arbitrary[[:space:]]Relations[[:space:]]from[[:space:]]Pretrained[[:space:]]Language[[:space:]]Models/0bb6a8fa-b40c-4f22-8940-95411b79bc73_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Better[[:space:]]Language[[:space:]]Models[[:space:]]of[[:space:]]Code[[:space:]]through[[:space:]]Self-Improvement/3f3acf36-4d5c-402e-88db-fa3521de22ad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Better[[:space:]]Zero-Shot[[:space:]]Reasoning[[:space:]]with[[:space:]]Self-Adaptive[[:space:]]Prompting/78205101-7057-4f07-9351-7773e5928f9e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Beyond[[:space:]]Positive[[:space:]]Scaling_[[:space:]]How[[:space:]]Negation[[:space:]]Impacts[[:space:]]Scaling[[:space:]]Trends[[:space:]]of[[:space:]]Language[[:space:]]Models/2f38d52d-fb68-4d9d-b74c-57bcc8239ae7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Beyond[[:space:]]Triplet_[[:space:]]Leveraging[[:space:]]the[[:space:]]Most[[:space:]]Data[[:space:]]for[[:space:]]Multimodal[[:space:]]Machine[[:space:]]Translation/a31314b6-17db-41ed-9ad1-4ca051d5085b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Bi-level[[:space:]]Finetuning[[:space:]]with[[:space:]]Task-dependent[[:space:]]Similarity[[:space:]]Structure[[:space:]]for[[:space:]]Low-resource[[:space:]]Training/88fa5869-4ead-43cd-a412-c7f83e6ce708_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Bias[[:space:]]Beyond[[:space:]]English_[[:space:]]Counterfactual[[:space:]]Tests[[:space:]]for[[:space:]]Bias[[:space:]]in[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]in[[:space:]]Four[[:space:]]Languages/250fd2fe-bf85-4229-81df-443c8c2ead89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Bidirectional[[:space:]]Transformer[[:space:]]Reranker[[:space:]]for[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction/298318f2-7f9a-4880-8d1c-5d7b87e9cc1f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/BigVideo_[[:space:]]A[[:space:]]Large-scale[[:space:]]Video[[:space:]]Subtitle[[:space:]]Translation[[:space:]]Dataset[[:space:]]for[[:space:]]Multimodal[[:space:]]Machine[[:space:]]Translation/8e43b454-6dac-4ae4-ae8b-7a1a3668ca1f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Boost[[:space:]]Transformer-based[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]GPU-Friendly[[:space:]]Sparsity[[:space:]]and[[:space:]]Quantization/051f6dc0-5993-4b54-89e6-2ccd2a9db3d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Boosting[[:space:]]Distress[[:space:]]Support[[:space:]]Dialogue[[:space:]]Responses[[:space:]]with[[:space:]]Motivational[[:space:]]Interviewing[[:space:]]Strategy/f21c65d5-325d-4c0d-a346-78b7c29673b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Boosting[[:space:]]Event[[:space:]]Extraction[[:space:]]with[[:space:]]Denoised[[:space:]]Structure-to-Text[[:space:]]Augmentation/890bc167-f220-4d7b-a18e-e79fd295cc6d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Boosting[[:space:]]Text[[:space:]]Augmentation[[:space:]]via[[:space:]]Hybrid[[:space:]]Instance[[:space:]]Filtering[[:space:]]Framework/e810a45f-a7e6-4ce6-9568-5eb41c97d795_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Boosting[[:space:]]Zero-shot[[:space:]]Cross-lingual[[:space:]]Retrieval[[:space:]]by[[:space:]]Training[[:space:]]on[[:space:]]Artificially[[:space:]]Code-Switched[[:space:]]Data/4990fc56-ddc4-4567-ad3a-7fb8257cbf3f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Bridge[[:space:]]the[[:space:]]Gap[[:space:]]Between[[:space:]]CV[[:space:]]and[[:space:]]NLP![[:space:]]A[[:space:]]Gradient-based[[:space:]]Textual[[:space:]]Adversarial[[:space:]]Attack[[:space:]]Framework/2f998aad-dc22-4c6e-a354-a4d9016ce4b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Bridging[[:space:]]the[[:space:]]Granularity[[:space:]]Gap[[:space:]]for[[:space:]]Acoustic[[:space:]]Modeling/3592a72e-bfba-437d-9a57-57f42f3c9374_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/C-XNLI_[[:space:]]Croatian[[:space:]]Extension[[:space:]]of[[:space:]]XNLI[[:space:]]Dataset/89a8e5d3-fa43-4cab-acf8-a149bccc7064_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CDA_[[:space:]]A[[:space:]]Contrastive[[:space:]]Data[[:space:]]Augmentation[[:space:]]Method[[:space:]]for[[:space:]]Alzheimer’s[[:space:]]Disease[[:space:]]Detection/dc0a73ce-7b3c-4548-850b-3fc4178c61d0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CFL_[[:space:]]Causally[[:space:]]Fair[[:space:]]Language[[:space:]]Models[[:space:]]Through[[:space:]]Token-level[[:space:]]Attribute[[:space:]]Controlled[[:space:]]Generation/2580b0cd-d343-42d4-b0a0-fff19849a1ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CIF-PT_[[:space:]]Bridging[[:space:]]Speech[[:space:]]and[[:space:]]Text[[:space:]]Representations[[:space:]]for[[:space:]]Spoken[[:space:]]Language[[:space:]]Understanding[[:space:]]via[[:space:]]Continuous[[:space:]]Integrate-and-Fire[[:space:]]Pre-Training/5b364f23-a30e-4a63-8fa7-f386b2d481dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CKDST_[[:space:]]Comprehensively[[:space:]]and[[:space:]]Effectively[[:space:]]Distill[[:space:]]Knowledge[[:space:]]from[[:space:]]Machine[[:space:]]Translation[[:space:]]to[[:space:]]End-to-End[[:space:]]Speech[[:space:]]Translation/dd472cda-57ba-4acc-ae84-ddf1f0abcd8e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CLIPText_[[:space:]]A[[:space:]]New[[:space:]]Paradigm[[:space:]]for[[:space:]]Zero-shot[[:space:]]Text[[:space:]]Classification/3e37f87f-d076-4276-aa1b-4d63541f20c2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/COBRA[[:space:]]Frames_[[:space:]]Contextual[[:space:]]Reasoning[[:space:]]about[[:space:]]Effects[[:space:]]and[[:space:]]Harms[[:space:]]of[[:space:]]Offensive[[:space:]]Statements/9ccb742a-19b9-4b88-b702-444c14e42d26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/COCKATIEL_[[:space:]]COntinuous[[:space:]]Concept[[:space:]]ranKed[[:space:]]ATtribution[[:space:]]with[[:space:]]Interpretable[[:space:]]ELements[[:space:]]for[[:space:]]explaining[[:space:]]neural[[:space:]]net[[:space:]]classifiers[[:space:]]on[[:space:]]NLP/40ea06fc-5506-42f5-b997-a7f63a71b447_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CSS_[[:space:]]A[[:space:]]Large-scale[[:space:]]Cross-schema[[:space:]]Chinese[[:space:]]Text-to-SQL[[:space:]]Medical[[:space:]]Dataset/96628818-d3d2-4176-9368-479e0a9498c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CaPE_[[:space:]]Contrastive[[:space:]]Parameter[[:space:]]Ensembling[[:space:]]for[[:space:]]Reducing[[:space:]]Hallucination[[:space:]]in[[:space:]]Abstractive[[:space:]]Summarization/b70d6452-fc0e-4ff3-a32b-43bc967ee1e6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Cross-Lingual[[:space:]]Transferability[[:space:]]of[[:space:]]Multilingual[[:space:]]Transformers[[:space:]]Be[[:space:]]Activated[[:space:]]Without[[:space:]]End-Task[[:space:]]Data_/04c12e62-2e5e-460f-80b4-a7e5e372cc21_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Diffusion[[:space:]]Model[[:space:]]Achieve[[:space:]]Better[[:space:]]Performance[[:space:]]in[[:space:]]Text[[:space:]]Generation[[:space:]]_[[:space:]]Bridging[[:space:]]the[[:space:]]Gap[[:space:]]between[[:space:]]Training[[:space:]]and[[:space:]]Inference[[:space:]]!/1cf2c517-cee2-4039-be92-8f652c3aa013_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Language[[:space:]]Models[[:space:]]Be[[:space:]]Specific_[[:space:]]How_/11666e62-a18e-4fd8-86be-5f95f806edd5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Categorial[[:space:]]grammar[[:space:]]induction[[:space:]]from[[:space:]]raw[[:space:]]data/a86b65ba-dad1-4cdf-8082-5f9c61ff698a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Causal[[:space:]]Intervention[[:space:]]for[[:space:]]Mitigating[[:space:]]Name[[:space:]]Bias[[:space:]]in[[:space:]]Machine[[:space:]]Reading[[:space:]]Comprehension/46ac0eb9-d16a-4830-ab91-d0c3a0b665bb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Causal[[:space:]]Matching[[:space:]]with[[:space:]]Text[[:space:]]Embeddings_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]in[[:space:]]Estimating[[:space:]]the[[:space:]]Causal[[:space:]]Effects[[:space:]]of[[:space:]]Peer[[:space:]]Review[[:space:]]Policies/e44742c5-8486-4d4b-a0f1-11ef409931e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Causal[[:space:]]interventions[[:space:]]expose[[:space:]]implicit[[:space:]]situation[[:space:]]models[[:space:]]for[[:space:]]commonsense[[:space:]]language[[:space:]]understanding/b28bc1fb-510d-43a1-80ed-c384eaf5e03c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CausalDialogue_[[:space:]]Modeling[[:space:]]Utterance-level[[:space:]]Causality[[:space:]]in[[:space:]]Conversations/0193e320-2f62-4770-a54f-170d2f2ea127_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Chain[[:space:]]of[[:space:]]Thought[[:space:]]Prompting[[:space:]]Elicits[[:space:]]Knowledge[[:space:]]Augmentation/f2164c55-344e-450f-aa2f-ca15b22eb2ef_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Challenging[[:space:]]BIG-Bench[[:space:]]Tasks[[:space:]]and[[:space:]]Whether[[:space:]]Chain-of-Thought[[:space:]]Can[[:space:]]Solve[[:space:]]Them/4fe76542-c100-4864-b7aa-83c145d8d459_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Character[[:space:]]Coreference[[:space:]]Resolution[[:space:]]in[[:space:]]Movie[[:space:]]Screenplays/d7de8d22-ced4-4717-bbf7-0581e70bc627_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Characterizing[[:space:]]the[[:space:]]Impacts[[:space:]]of[[:space:]]Instances[[:space:]]on[[:space:]]Robustness/cc0d02cd-2829-4bb0-a7a3-5ddbea2b3bde_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Check-COVID_[[:space:]]Fact-Checking[[:space:]]COVID-19[[:space:]]News[[:space:]]Claims[[:space:]]with[[:space:]]Scientific[[:space:]]Evidence/95c00212-34bf-4dae-9624-596be9689568_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Claim-Dissector_[[:space:]]An[[:space:]]Interpretable[[:space:]]Fact-Checking[[:space:]]System[[:space:]]with[[:space:]]Joint[[:space:]]Re-ranking[[:space:]]and[[:space:]]Veracity[[:space:]]Prediction/740d87df-c1bb-4f26-afda-60412ea1a703_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ClaimDiff_[[:space:]]Comparing[[:space:]]and[[:space:]]Contrasting[[:space:]]Claims[[:space:]]on[[:space:]]Contentious[[:space:]]Issues/8c0340a8-f0c5-4172-8831-87c221254b8d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Class[[:space:]]Lifelong[[:space:]]Learning[[:space:]]for[[:space:]]Intent[[:space:]]Detection[[:space:]]via[[:space:]]Structure[[:space:]]Consolidation[[:space:]]Networks/4ff876cf-ee47-485f-9bff-eb85c268fa46_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Class-Adaptive[[:space:]]Self-Training[[:space:]]for[[:space:]]Relation[[:space:]]Extraction[[:space:]]with[[:space:]]Incompletely[[:space:]]Annotated[[:space:]]Training[[:space:]]Data/ea255dfd-3f40-4f9b-a975-d9e5bbed0c86_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Click_[[:space:]]Controllable[[:space:]]Text[[:space:]]Generation[[:space:]]with[[:space:]]Sequence[[:space:]]Likelihood[[:space:]]Contrastive[[:space:]]Learning/2c3e2f31-3d5e-41f1-9026-7f296e9184ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Client-Customized[[:space:]]Adaptation[[:space:]]for[[:space:]]Parameter-Efficient[[:space:]]Federated[[:space:]]Learning/b6fd3ae3-d337-4709-b1b6-c3c9ab4291cd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Clustering-Aware[[:space:]]Negative[[:space:]]Sampling[[:space:]]for[[:space:]]Unsupervised[[:space:]]Sentence[[:space:]]Representation/2b1bb59c-f44b-4bbe-8c18-81776afb252e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoAug_[[:space:]]Combining[[:space:]]Augmentation[[:space:]]of[[:space:]]Labels[[:space:]]and[[:space:]]Labelling[[:space:]]Rules/a274c196-a5ab-436b-b8b9-12e6fa1704b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoMave_[[:space:]]Contrastive[[:space:]]Pre-training[[:space:]]with[[:space:]]Multi-scale[[:space:]]Masking[[:space:]]for[[:space:]]Attribute[[:space:]]Value[[:space:]]Extraction/dc7458c8-0c7c-4461-80a0-703f5aaf0447_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoMix_[[:space:]]Guide[[:space:]]Transformers[[:space:]]to[[:space:]]Code-Mix[[:space:]]using[[:space:]]POS[[:space:]]structure[[:space:]]and[[:space:]]Phonetics/ade88a51-a2f4-437d-9b99-c0b35a7b8046_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoRRPUS_[[:space:]]Code-based[[:space:]]Structured[[:space:]]Prompting[[:space:]]for[[:space:]]Neurosymbolic[[:space:]]Story[[:space:]]Understanding/71512c2a-8376-4e9e-aa84-ca69a8131592_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Coarse-to-fine[[:space:]]Few-shot[[:space:]]Learning[[:space:]]for[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition/023adf43-4de8-4cec-a35b-fc36c016008b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Code[[:space:]]Execution[[:space:]]with[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models/f84bee97-1570-4171-9ef8-11cbb58b4075_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Code-Switched[[:space:]]Text[[:space:]]Synthesis[[:space:]]in[[:space:]]Unseen[[:space:]]Language[[:space:]]Pairs/b627b8b3-6d9e-49f2-a9f8-3a27be6b8015_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CodePrompt_[[:space:]]Task-Agnostic[[:space:]]Prefix[[:space:]]Tuning[[:space:]]for[[:space:]]Program[[:space:]]and[[:space:]]Language[[:space:]]Generation/7f6aa9d3-5907-4e0f-8830-86a860ebc33b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Coherent[[:space:]]or[[:space:]]Not_[[:space:]]Stressing[[:space:]]a[[:space:]]Neural[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Discourse[[:space:]]Coherence[[:space:]]in[[:space:]]Multiple[[:space:]]Languages/30e11dca-164c-4670-973d-dc3881674ea2_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Automatic Named Entity Obfuscation in Speech/f423cca1-1aeb-43bd-8d25-5845a22e4c4c_content_list.json

@@ -0,0 +1,1361 @@
+[
+    {
+        "type": "text",
+        "text": "Automatic Named Entity Obfuscation in Speech",
+        "text_level": 1,
+        "bbox": [
+            248,
+            90,
+            746,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Judita Preiss  \nUniversity of Sheffield, Information School  \nThe Wave, 2 Whitham Road  \nSheffield S10 2AH  \njudita.preiss@sheffield.ac.uk",
+        "bbox": [
+            324,
+            131,
+            680,
+            214
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sharing data containing personal information often requires its anonymization, even when consent for sharing was obtained from the data originator. While approaches exist for automated anonymization of text, the area is not as thoroughly explored in speech. This work focuses on identifying, replacing and inserting replacement named entities synthesized using voice cloning into original audio thereby retaining prosodic information while reducing the likelihood of deanonymization. The approach employs a novel named entity recognition (NER) system built directly on speech by training HuBERT (Hsu et al., 2021) using the English speech NER dataset (Yadav et al., 2020). Name substitutes are found using a masked language model and are synthesized using text to speech voice cloning (Eren and Team, 2021), upon which the substitute named entities are re-inserted into the original text. The approach is prototyped on a sample of the LibriSpeech corpus (Panayotov et al., 2015) with each step evaluated individually.",
+        "bbox": [
+            144,
+            280,
+            460,
+            607
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            620,
+            260,
+            634
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Privacy concerns, particularly where an individual could be identified, preclude sharing and therefore automatic exploitation of many data sources. Anonymization, the removal of identifying information, has been automated for text (Lison et al., 2021), including large scale applications such as in clinical (Hartman et al., 2020) or legal settings (Oksanen et al., 2022), with off-the-shelf systems having reported performance of  $90 + \\%$  (Hartman et al., 2020). To minimize the risk of re-identification, obfuscation - replacing identifying information with a different substitute of the same type - has been explored as an alternative to replacing identifying information with a generic marker (Sousa and Kern, 2022). The main focus in speech has been on voice anonymization, which may not be a problem with speaker consent, with the removal of",
+        "bbox": [
+            112,
+            645,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "identifying information receiving less attention. To our knowledge, this is the first prototype to perform named entity obfuscation directly, in the original speaker's voice. Aside from voice cloning, it explores a named entity recognition approach based directly on audio signal and uses language model masking to find appropriate substitutions.",
+        "bbox": [
+            507,
+            253,
+            884,
+            365
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in speech models, particularly the inclusion of language models within the speech model itself (e.g. HuBERT (Hsu et al., 2021)) gives models greater insight into expected contexts. Previous work on named entity recognition (NER) in speech frequently employs a two step approach, transcribing speech first, followed by the application of existing named entity techniques (Yadav et al., 2020). However, this process has the potential to compound errors as errors in transcription will increase the probability of error in NER. We suggest that the addition of language models into the speech model gives these sufficient power to perform NER directly, and therefore that transcribing (automatic speech recognition, ASR) and NER can be separated, and used to provide a confidence measure in their performance. Divided, the two do not propagate errors in the same way; in fact, treating ASR and NER separately allows one to fix (some of the) errors of the other. The proposed second (final) ASR pass merely produces a confidence value in the result to decide whether a manual check should be performed.",
+        "bbox": [
+            507,
+            368,
+            884,
+            737
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The success of few shot learning, where a limited number of examples is used to generalize a pre-trained deep learning model to a new situation, for text-to-speech – and specifically voice cloning (Zhang and Lin, 2022) – enables an alternative, equivalent but different, entity to be inserted in the audio signal in place of the original while preserving the prosody information throughout. While large databases of potential replacement entities can be used to select a substitution, these may not preserve necessary properties (such as gender). Al-",
+        "bbox": [
+            507,
+            741,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "615",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 615-622 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            236,
+            945,
+            759,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ternatively, word embeddings have been used to suggest close (in the multi-dimensional space) alternatives (Abdalla et al., 2020), however these can suffer from the same drawback. We propose using a more contextualized alternative to word embeddings, a masked language model (Devlin et al., 2019), where the model is trained by hiding (masking) words and predictions of the original word are made based on their context.",
+        "bbox": [
+            112,
+            84,
+            489,
+            227
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This work makes the following contributions: (1) a complete obfuscation pipeline for names in speech $^{1}$ , (2) a named entity recognizer built directly on speech without requiring text transcription first, (3) alternative (obfuscated) entity replacement selection via masking language model, and (4) confidence annotated system output, allowing for manual correction and / or selection of shareable instances. Section 2 contains the methodology with results in Section 3. Section 4 presents the conclusions and future work.",
+        "bbox": [
+            115,
+            230,
+            489,
+            405
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            422,
+            265,
+            437
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The steps of the overall pipeline, which takes in an audio file and produces an obfuscated audio file along with a confidence value, can be found in Figure 1. The approach comprises of three main parts: 1) identification of named entities (NEs) in the audio, 2) finding an equivalent alternative for the original NEs, and 3) reconstructing the original audio to incorporate the replacement NEs. The reconstructed audio can further be used to obtain a confidence value.",
+        "bbox": [
+            112,
+            449,
+            487,
+            609
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Identification of named entities",
+        "text_level": 1,
+        "bbox": [
+            112,
+            624,
+            405,
+            639
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To enable the direct use of a language model on speech input for the purpose of named entity recognition (NER), a dataset of audio recordings with annotated NEs is required. The English speech NER dataset (Yadav et al., 2020), which consists of 70,769 waveforms with transcripts annotated with person, location and organization NEs, is used for fine-tuning the Hidden-Unit BERT speech model (HuBERT) (Hsu et al., 2021). HuBERT was selected over other speech models since it learns both acoustic and language models from its inputs and therefore has an increased awareness of context. The success of language models on text NER has demonstrated how crucial context is for this",
+        "bbox": [
+            112,
+            646,
+            489,
+            870
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "task, and using a model which incorporates both an acoustic and a language model (over acoustic only) allows the approach to exploit the information used in text NER, while managing to avoid the need for a transcript.",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For training, NE annotations need to be converted to a suitable format, indicating the presence or absence of a NE in each position. Following the inside-outside(-beginning) chunking common to many NER approaches (Tjong Kim Sang and De Meulder, 2003), three formats were explored: 1) character level annotation, mapping each character to either  $o$  for a character outside of a named entity, space, or  $n$ ,  $l$ ,  $e$  for characters within person, location or organization entities respectively, 2) the same character level annotation with separate characters added to denote the beginning of each type of NE (mapping the sentence TELL JACK to oooo mnnn with  $m$  denoting the start of a person NE), 3) and, for completeness, annotation was also explored at word level.",
+        "bbox": [
+            507,
+            167,
+            884,
+            423
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "With the training parameters shown in Appendix A.1, the best NE performance was obtained from the first annotation approach, where NE beginnings were not explicitly annotated. The lower performance of the second annotation approach can be attributed to the low quantity of training data for the beginning marker annotations. While word level annotation was explored, it is likely to need a far greater quantity of data to enable mapping of different length inputs to a single label.",
+        "bbox": [
+            507,
+            426,
+            885,
+            586
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Separately, HuBERT was also fine-tuned for automatic speech recognition (ASR), i.e. for transcribing text from audio. Identical training data was used, with annotation being the transcription provided as part of the NE annotation (with NE annotation removed). The same parameters were employed for its training. Alongside the predicted (NE or ASR) annotation, prediction output also yields an offset which can be converted to a time offset. This can be used to identify the position of the NE(s) to be replaced, and after a greedy alignment of the two outputs, the original transcription of the original NE(s) can be extracted.",
+        "bbox": [
+            507,
+            589,
+            885,
+            797
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Finding an alternative NE",
+        "text_level": 1,
+        "bbox": [
+            507,
+            814,
+            761,
+            829
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Once a person NE is identified, a suitable equivalent substitution needs to be obtained, i.e. we want to find the word which could replace the NE in the text if the NE was hidden. This is precisely the concept behind masked language models (MLMs):",
+        "bbox": [
+            507,
+            839,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>The code for the obfuscation pipeline is available from: https://github.com/juditapreiss/speech_ne_obfuscation",
+        "bbox": [
+            112,
+            879,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "616",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/486fcf9d398058710632011d88c10eff46b67d5226d5edf62beb14bd3d9045d1.jpg",
+        "image_caption": [
+            "Figure 1: Obfuscation pipeline"
+        ],
+        "image_footnote": [
+            "Speech to obfuscate as a wav file",
+            "Transcribed text with NEs & timings",
+            "Obfuscated text",
+            "Generated text using cloned voice",
+            "Extracted timings of substituted NEs",
+            "Inserted generated NEs into orig text",
+            "WER and CER based confidence"
+        ],
+        "bbox": [
+            114,
+            80,
+            537,
+            373
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "these models learn their weights so that given a sentence with a hidden (masked) word, the model will output the complete original sentence. The (ASR extracted) original sentences with NEs (as identified by the NE tuned model) masked were passed to a MLM. Three MLM models were explored: BERT, bert-large-uncased model (Devlin et al., 2019), ALBERT, albert-xxlarge-v2, model (Lan et al., 2019) and the distilled RoBERTa base, distilroberta-base, model (Sanh et al., 2019). Each model, with no additional tuning, results in a (pre-specified) number of predictions for each NE in the sentence. Since the models used different datasets in training, their predictions are expected to be different: for example, some may suggest pronouns rather than names.",
+        "bbox": [
+            112,
+            423,
+            489,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given the propensity of the MLM to return substitutions which are not names (for example, for the sentence you should call Stella, the model returns you should call him, you should call them, you should call 911 etc), an external list of people names is used for the validation of the proposed suggestions<sup>2</sup> and the highest scoring substitution is returned. Heuristically, the original name is matched against the list to identify whether it is a first or a last name (where possible) and names of the same type suggested by the MLM are returned. Simple rules are employed (last of a sequence of",
+        "bbox": [
+            112,
+            682,
+            489,
+            875
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "names is a last name, a single name without a title is a first name etc) to decide on a substitution when the original name does not appear in either the first or last name list. Given the nature of MLMs, suggested alternatives are likely to be more common words: as a positive side effect, this should make them easier to render with voice cloning as they may already appear in the reference speech. Should MLM fail to propose any suitable substitutions, one is selected at random from the first & last name lists, subject to the same heuristic rules.",
+        "bbox": [
+            507,
+            423,
+            884,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Reconstruction of original audio",
+        "text_level": 1,
+        "bbox": [
+            507,
+            619,
+            808,
+            634
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, the substitute NE is to be re-inserted into the original audio. To reduce the risk of de-identification via the extraction of entities which failed to be identified and therefore stayed in their original form, the substitute entity needs to be produced in the speaker's voice. The YourTTS (Casanova et al., 2021) model, which offers the ability for fine-tuning with less than one minute of speech while achieving good results with reasonable quality, can be used to generate the substitute sentence with all available speech of the speaker provided as reference. Note that it is not necessary to remove the original sentence from the reference data: in fact, its presence may result in more accurate rendering of the substitute sentence. The pre-trained model used in this work (tts_models/multilingual/multi-dataset/your_tts)",
+        "bbox": [
+            505,
+            645,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "2In this work, https://github.com/dominictarr/random-name are used.",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "617",
+        "bbox": [
+            484,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "was trained on the the voice cloning toolkit (VCTK) dataset (Yamagishi et al., 2019) which contains approximately 400 sentence, selected from newspaper text, uttered by 108-110 different speakers, giving it its generalization power. Aside from the reference passed to the model on the command line, no tuning or training of the YourTTS model is done in this work.",
+        "bbox": [
+            112,
+            84,
+            487,
+            211
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The ASR transcribed text with the substituted NE is generated, rather than the substitution alone, to ensure that the intonation as closely matches the substitution's position in the sentence. The average amplitude of the generated audio is matched to that of the original segment using the Python pydub library. The generated audio is again passed through the HuBERT based NE recognizer, to identify the location of the substituted NE in the generated audio and allow its extraction (note that in this pass, it is not necessary to perform ASR - only the offsets of the replacement NE are required). Should the NE recognizer not identify the same number of NEs as were present in the original, the instance is flagged for manual review.",
+        "bbox": [
+            115,
+            212,
+            489,
+            453
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each NE in the text, a pair of start and end offsets are available: one pair extracted by the HubERT based NE extraction from the original audio and a second pair from the audio generated from the substituted text. This allows the new NEs to be inserted in place of the original NEs. The splicing and concatenation of the waveforms is also performed using the pydub library.",
+        "bbox": [
+            112,
+            455,
+            489,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A second HuBERT based ASR pass over the newly constructed (substituted) audio, and its comparison against the substituted text using word error rate (WER) and character error rate (CER) gives measures of confidence. Both the metrics, commonly used for evaluation of ASR, allow for sequences of different length to the target – the further the reconstructed audio is from the target sentence, the less likely it is that the substitution will go unnoticed.",
+        "bbox": [
+            112,
+            585,
+            489,
+            744
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Results and discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            758,
+            341,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Identification of named entities",
+        "text_level": 1,
+        "bbox": [
+            112,
+            785,
+            403,
+            799
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The 70,769 training corpus, sampled at  $16\\mathrm{kHz}$ , is divided up into  $70\\%$  for training (49,540 instances), and  $15\\%$  for both validation and evaluation (10,615 examples). The hubert-base-1s960 model is used with parameters listed in Appendix A.1. The performance in training, indicated via WER and CER, is shown in Table 1 for both ASR and NER.",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/2e2c2154bd75ecf1e8b4e8c2c4b40209d9fc0e6e95660a663d6fd99ce7199f92.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Eval WER</td><td>Eval CER</td></tr><tr><td>ASR</td><td>0.142</td><td>-</td></tr><tr><td>NE</td><td>0.199</td><td>0.063</td></tr></table>",
+        "bbox": [
+            509,
+            82,
+            764,
+            130
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/6b2c6c4e139925d7de084f6ce80f57b8bc2548533026695148045bad40ac8c6b.jpg",
+        "table_caption": [
+            "Table 1: Metric results of the ASR and NE HuBERT based models"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>MLM</td><td>Avg ASR</td><td>NE</td><td>Avg confidence</td></tr><tr><td>ALBERT</td><td>0.980</td><td>13/20</td><td>0.109</td></tr><tr><td>BERT</td><td>0.980</td><td>13/20</td><td>0.098</td></tr><tr><td>RoBERTa</td><td>0.980</td><td>13/20</td><td>0.106</td></tr></table>",
+        "bbox": [
+            510,
+            186,
+            894,
+            254
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Evaluation of individual steps",
+        "bbox": [
+            561,
+            263,
+            826,
+            277
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the purpose of the demonstrating the viability of the prototype, no hyperparameter optimization was performed, and the larger HuBERT models were not employed, however improvement in performance of both models are expected should this be pursued.",
+        "bbox": [
+            507,
+            304,
+            882,
+            400
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Finding an alternative NE",
+        "text_level": 1,
+        "bbox": [
+            507,
+            413,
+            759,
+            428
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A small scale evaluation is performed on a sample of 20 sentences selected at random from the LibriSpeech corpus (Panayotov et al., 2015) across 6 speakers. Sentence selection was subject to them containing a person named entity. While detailed results for the individual steps can be found in Table 2, it should be noted that – for the purposes of this work – the focus is the accuracy of the extraction of the correct NE. The stated accuracy is therefore somewhat misleading: in a number of cases, such as the word Raphael, the named entity is divided into two separate words, suggesting two consecutive named entities. However, this issue is corrected when the NE output is aligned with ASR output and the two separate NE instances are (correctly) merged. Cases with NEs which cannot be aligned are flagged up for manual intervention. The average ASR and (exact match) NE identification do not vary when a different MLM is employed, as this only effects the selection of the substituted name, resulting in different average confidence values.",
+        "bbox": [
+            507,
+            435,
+            884,
+            785
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Reconstruction of original audio",
+        "text_level": 1,
+        "bbox": [
+            507,
+            801,
+            808,
+            816
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The voice cloning model requires some reference audio for the speaker: for the 6 selected speakers, 4 have less than 5 audio files (two having 3, and one having only 2 files) in the dataset. The quantity of data used as reference is likely to impact the quality (in terms of its similarity to the original speaker)",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "618",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "of the generated text. Given the likely scenarios of deployment, such as dialogues where more than 2 sentences of speech per speaker are available, this may not be representative of the results obtainable with the pipeline. However, it should be noted that even if all substituted instances can be identified as substitutions, the system is equal to a masking technique (where an entity is replaced with a fixed entity, such as a bleep).",
+        "bbox": [
+            112,
+            84,
+            489,
+            231
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            242,
+            247,
+            256
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The prototype described shows the steps of an obfuscation pipeline for speech, which results in substituted person named entities uttered in the original speakers voice and replaced in the original audio signal. The prototype makes use of a named entity recognizer built directly on top of audio input, and employs masked language models to generate the substituted entity. It offers an end-to-end automatic solution enabling the sharing of speech with identifying information removed.",
+        "bbox": [
+            112,
+            268,
+            489,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The resulting obfuscated speech remains in the original speaker's voice, allowing for the application of traditional speaker anonymization approaches to mask the speaker's identity. The original prosody can be protected by applying a transformation such as waveform change, offering a significant advantage over a technique which generates a complete obfuscated transcription (instead of splicing an obfuscated entity into original speech).",
+        "bbox": [
+            112,
+            429,
+            489,
+            576
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            586,
+            220,
+            602
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The cloning model used, YourTTS, is trained on the VCTK dataset which consists of high-quality speech signal. It is therefore unclear whether the same accuracy would be obtained with lower quality signal which may contain some background noise. (However, it should again be noted that even if all substituted instances are identifiable in the output, the system is equivalent to a masking model.)",
+        "bbox": [
+            112,
+            612,
+            489,
+            757
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The selection of a person NE replacement does not currently account for continuity: if the same person entity is referred to later, it may be substituted with a different entity to the previous occasion. In addition, the back-off strategy ignores aspects such as gender.",
+        "bbox": [
+            112,
+            758,
+            489,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To show the approach feasible, very little optimization was performed. Further training and parameter optimization is likely to lead to improved performance for both ASR and NER models.",
+        "bbox": [
+            112,
+            854,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The approach is currently only implemented for person NEs but it could be extended very simply to other types of NEs. However, the degree to which other entity types require obfuscation in speech is not clear to us as mentions of organizations may well not be identifying at all.",
+        "bbox": [
+            507,
+            84,
+            884,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            197,
+            660,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Aside from the ethical concerns regarding voice cloning (covered in e.g. YourTTS (Casanova et al., 2021)), deployment would require a detailed evaluation of risk of de-identification. It is believed that the final confidence and the accuracy of each step can be combined to significantly reduce this risk. The voice itself also offers options for identification: the value of yielding substitutions in the original speaker's voice (and keeping the original prosody) would need to be weighed up against approaches which anonymize voice but preserve prosodic information.",
+        "bbox": [
+            507,
+            225,
+            884,
+            419
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            449,
+            608,
+            464
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mohamed Abdalla, Moustafa Abdalla, Frank Rudzicz, and Graeme Hirsto. 2020. Using word embeddings to improve the privacy of clinical notes. J Am Med Inform Assoc, 27(6):901-907.",
+            "Edresson Casanova, Julian Weber, Christopher Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir Antonelli Ponti. 2021. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. CoRR, abs/2112.02418.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT.",
+            "Gölge Eren and The Coqui TTS Team. 2021. Coqui TTS. https://github.com/coqui-ai/TTS.",
+            "Tzvika Hartman, Michael D. Howell, Jeff Dean, Shlomo Hoory, Ronit Slyper, Itay Laish, Oren Gilon, Danny Vainstein, Greg Corrado, Katherine Chou, Ming Jack Po, Jutta Williams, Scott Ellis, Gavin Bee, Avinatan Hassidim, Rony Amira, Genady Beryozkin, Idan Szpektor, and Yossi Matias. 2020. Customization scenarios for de-identification of clinical notes. BMC Med Inform Decis Mak, 20(14).",
+            "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. CoRR, abs/2106.07447."
+        ],
+        "bbox": [
+            509,
+            474,
+            885,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "619",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. ALBERT: A lite BERT for self-supervised learning of language representations. CoRR, abs/1909.11942.",
+            "Pierre Lison, Ildiko Pilán, David Sanchez, Montserrat Batet, and Lilja Øvrelid. 2021. Anonymisation models for text data: State of the art, challenges and future directions. 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 4188-4203, Online. Association for Computational Linguistics.",
+            "Arttu Oksanen, Eero Hyvönen, Minna Tamper, Jouni Tuominen, Henna Ylimaa, Katja Löytynoja, Matti Kokkonen, and Aki Hietanen. 2022. An anonymization tool for open data publication of legal documents. In Joint Proceedings of the 3th International Workshop on Artificial Intelligence Technologies for Legal Documents (AI4LEGAL 2022) and the 1st International Workshop on Knowledge Graph Summarization (KGSum 2022) co-located with the 21st International Semantic Web Conference (ISWC 2022), Virtual Event, Hangzhou, China, October 23-24, 2022, volume 3257 of CEUR Workshop Proceedings, pages 12-21. CEUR-WS.org.",
+            "Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an ASR corpus based on public domain audio books. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE.",
+            "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108.",
+            "Samuel Sousa and Roman Kern. 2022. How to keep text private? a systematic review of deep learning methods for privacy-preserving natural language processing. Artif Intell Rev.",
+            "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147.",
+            "Hemant Yadav, Sreyan Ghosh, Yi Yu, and Rajiv Ratn Shah. 2020. End-to-end named entity recognition from English speech. CoRR, abs/2005.11184.",
+            "Junichi Yamagishi, Christophe Veaux, and Kirsten MacDonald. 2019. CSTR VCTK corpus: English multispeaker corpus for CSTR voice cloning toolkit (version 0.92).",
+            "Haitong Zhang and Yue Lin. 2022. Improve few-shot voice cloning using multi-modal learning. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 8317-8321."
+        ],
+        "bbox": [
+            114,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A Model training details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            741,
+            99
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A.1 HuBERT parameters",
+        "text_level": 1,
+        "bbox": [
+            509,
+            109,
+            729,
+            124
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "HubertForCTC hubert-base-1s960 model was used with frozen feature encoder, training size of 49540 examples, and validation and evaluation size of 10615, alongside the following parameters:",
+        "bbox": [
+            507,
+            130,
+            884,
+            193
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/272b82f2902f3ec259fa74ede7c33c27d2e26f1b440a6e87210536b0e251be33.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Parameter</td><td>Value</td></tr><tr><td>group_by_length</td><td>True</td></tr><tr><td>per_device_train_batch_size</td><td>8</td></tr><tr><td>per_device_eval_batch_size</td><td>batch_size</td></tr><tr><td>evaluation_strategy</td><td>&quot;steps&quot;</td></tr><tr><td>num_train_epochs</td><td>num_epochs</td></tr><tr><td>fp16</td><td>True</td></tr><tr><td>gradient_checkpointing</td><td>True</td></tr><tr><td>save_steps</td><td>500</td></tr><tr><td>eval_steps</td><td>500</td></tr><tr><td>learning_rate</td><td>1e-4</td></tr><tr><td>weight Decay</td><td>0.005</td></tr><tr><td>warmup_steps</td><td>1000</td></tr><tr><td>num_epochs</td><td>30</td></tr></table>",
+        "bbox": [
+            529,
+            196,
+            875,
+            423
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "620",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            143
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Unnumbered, follows conclusion",
+        "bbox": [
+            149,
+            143,
+            396,
+            158
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            129,
+            168,
+            552,
+            186
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement discusses potential risks of the work. This follows \"Limitations\".",
+        "bbox": [
+            149,
+            187,
+            752,
+            200
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            212,
+            695,
+            228
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Section 1",
+        "bbox": [
+            151,
+            230,
+            221,
+            243
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            256,
+            668,
+            272
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            213,
+            336
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            127,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            363,
+            248,
+            379
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            127,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            151,
+            407,
+            248,
+            422
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            127,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            151,
+            498,
+            248,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            127,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            151,
+            574,
+            248,
+            588
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            127,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            151,
+            633,
+            248,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            127,
+            658,
+            880,
+            739
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            151,
+            740,
+            248,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            494,
+            781
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sections 2 & 3",
+        "bbox": [
+            132,
+            785,
+            245,
+            800
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            812,
+            880,
+            844
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Appendix A1. The infrastructure was chosen by the HPC and therefore wasn't known - if needed, the experiments can be rerun with this constrained.",
+        "bbox": [
+            149,
+            846,
+            880,
+            876
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            884,
+            877,
+            908
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "621",
+        "bbox": [
+            485,
+            928,
+            512,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            129,
+            83,
+            878,
+            115
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experimental setup discussed, hyperparameter optimization not performed as stated as the paper only checks viability.",
+        "bbox": [
+            147,
+            117,
+            880,
+            148
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            129,
+            158,
+            880,
+            206
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Average used, clearly denoted. Section 3.",
+        "bbox": [
+            147,
+            208,
+            455,
+            223
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            127,
+            234,
+            880,
+            281
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            282,
+            349,
+            298
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            309,
+            877,
+            326
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            330,
+            215,
+            346
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            357,
+            880,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            390,
+            248,
+            405
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            416,
+            880,
+            463
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            464,
+            248,
+            480
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            491,
+            880,
+            538
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            539,
+            248,
+            555
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            565,
+            873,
+            581
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            583,
+            248,
+            598
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            609,
+            880,
+            639
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            640,
+            248,
+            656
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "622",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 7
+    }
+]

2023/Automatic Named Entity Obfuscation in Speech/f423cca1-1aeb-43bd-8d25-5845a22e4c4c_model.json

@@ -0,0 +1,1591 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.25,
+                0.091,
+                0.747,
+                0.112
+            ],
+            "angle": 0,
+            "content": "Automatic Named Entity Obfuscation in Speech"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.325,
+                0.132,
+                0.681,
+                0.215
+            ],
+            "angle": 0,
+            "content": "Judita Preiss  \nUniversity of Sheffield, Information School  \nThe Wave, 2 Whitham Road  \nSheffield S10 2AH  \njudita.preiss@sheffield.ac.uk"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.281,
+                0.461,
+                0.608
+            ],
+            "angle": 0,
+            "content": "Sharing data containing personal information often requires its anonymization, even when consent for sharing was obtained from the data originator. While approaches exist for automated anonymization of text, the area is not as thoroughly explored in speech. This work focuses on identifying, replacing and inserting replacement named entities synthesized using voice cloning into original audio thereby retaining prosodic information while reducing the likelihood of deanonymization. The approach employs a novel named entity recognition (NER) system built directly on speech by training HuBERT (Hsu et al., 2021) using the English speech NER dataset (Yadav et al., 2020). Name substitutes are found using a masked language model and are synthesized using text to speech voice cloning (Eren and Team, 2021), upon which the substitute named entities are re-inserted into the original text. The approach is prototyped on a sample of the LibriSpeech corpus (Panayotov et al., 2015) with each step evaluated individually."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.621,
+                0.262,
+                0.636
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.646,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Privacy concerns, particularly where an individual could be identified, preclude sharing and therefore automatic exploitation of many data sources. Anonymization, the removal of identifying information, has been automated for text (Lison et al., 2021), including large scale applications such as in clinical (Hartman et al., 2020) or legal settings (Oksanen et al., 2022), with off-the-shelf systems having reported performance of \\(90 + \\%\\) (Hartman et al., 2020). To minimize the risk of re-identification, obfuscation - replacing identifying information with a different substitute of the same type - has been explored as an alternative to replacing identifying information with a generic marker (Sousa and Kern, 2022). The main focus in speech has been on voice anonymization, which may not be a problem with speaker consent, with the removal of"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.885,
+                0.366
+            ],
+            "angle": 0,
+            "content": "identifying information receiving less attention. To our knowledge, this is the first prototype to perform named entity obfuscation directly, in the original speaker's voice. Aside from voice cloning, it explores a named entity recognition approach based directly on audio signal and uses language model masking to find appropriate substitutions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.369,
+                0.885,
+                0.738
+            ],
+            "angle": 0,
+            "content": "Recent advances in speech models, particularly the inclusion of language models within the speech model itself (e.g. HuBERT (Hsu et al., 2021)) gives models greater insight into expected contexts. Previous work on named entity recognition (NER) in speech frequently employs a two step approach, transcribing speech first, followed by the application of existing named entity techniques (Yadav et al., 2020). However, this process has the potential to compound errors as errors in transcription will increase the probability of error in NER. We suggest that the addition of language models into the speech model gives these sufficient power to perform NER directly, and therefore that transcribing (automatic speech recognition, ASR) and NER can be separated, and used to provide a confidence measure in their performance. Divided, the two do not propagate errors in the same way; in fact, treating ASR and NER separately allows one to fix (some of the) errors of the other. The proposed second (final) ASR pass merely produces a confidence value in the result to decide whether a manual check should be performed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The success of few shot learning, where a limited number of examples is used to generalize a pre-trained deep learning model to a new situation, for text-to-speech – and specifically voice cloning (Zhang and Lin, 2022) – enables an alternative, equivalent but different, entity to be inserted in the audio signal in place of the original while preserving the prosody information throughout. While large databases of potential replacement entities can be used to select a substitution, these may not preserve necessary properties (such as gender). Al-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "615"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.237,
+                0.946,
+                0.761,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 615-622 July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.228
+            ],
+            "angle": 0,
+            "content": "ternatively, word embeddings have been used to suggest close (in the multi-dimensional space) alternatives (Abdalla et al., 2020), however these can suffer from the same drawback. We propose using a more contextualized alternative to word embeddings, a masked language model (Devlin et al., 2019), where the model is trained by hiding (masking) words and predictions of the original word are made based on their context."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.231,
+                0.49,
+                0.406
+            ],
+            "angle": 0,
+            "content": "This work makes the following contributions: (1) a complete obfuscation pipeline for names in speech\\(^{1}\\), (2) a named entity recognizer built directly on speech without requiring text transcription first, (3) alternative (obfuscated) entity replacement selection via masking language model, and (4) confidence annotated system output, allowing for manual correction and / or selection of shareable instances. Section 2 contains the methodology with results in Section 3. Section 4 presents the conclusions and future work."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.423,
+                0.266,
+                0.438
+            ],
+            "angle": 0,
+            "content": "2 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.45,
+                0.488,
+                0.61
+            ],
+            "angle": 0,
+            "content": "The steps of the overall pipeline, which takes in an audio file and produces an obfuscated audio file along with a confidence value, can be found in Figure 1. The approach comprises of three main parts: 1) identification of named entities (NEs) in the audio, 2) finding an equivalent alternative for the original NEs, and 3) reconstructing the original audio to incorporate the replacement NEs. The reconstructed audio can further be used to obtain a confidence value."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.625,
+                0.406,
+                0.64
+            ],
+            "angle": 0,
+            "content": "2.1 Identification of named entities"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.647,
+                0.49,
+                0.871
+            ],
+            "angle": 0,
+            "content": "To enable the direct use of a language model on speech input for the purpose of named entity recognition (NER), a dataset of audio recordings with annotated NEs is required. The English speech NER dataset (Yadav et al., 2020), which consists of 70,769 waveforms with transcripts annotated with person, location and organization NEs, is used for fine-tuning the Hidden-Unit BERT speech model (HuBERT) (Hsu et al., 2021). HuBERT was selected over other speech models since it learns both acoustic and language models from its inputs and therefore has an increased awareness of context. The success of language models on text NER has demonstrated how crucial context is for this"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.166
+            ],
+            "angle": 0,
+            "content": "task, and using a model which incorporates both an acoustic and a language model (over acoustic only) allows the approach to exploit the information used in text NER, while managing to avoid the need for a transcript."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.168,
+                0.885,
+                0.424
+            ],
+            "angle": 0,
+            "content": "For training, NE annotations need to be converted to a suitable format, indicating the presence or absence of a NE in each position. Following the inside-outside(-beginning) chunking common to many NER approaches (Tjong Kim Sang and De Meulder, 2003), three formats were explored: 1) character level annotation, mapping each character to either \\(o\\) for a character outside of a named entity, space, or \\(n\\), \\(l\\), \\(e\\) for characters within person, location or organization entities respectively, 2) the same character level annotation with separate characters added to denote the beginning of each type of NE (mapping the sentence TELL JACK to oooo mnnn with \\(m\\) denoting the start of a person NE), 3) and, for completeness, annotation was also explored at word level."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.427,
+                0.887,
+                0.587
+            ],
+            "angle": 0,
+            "content": "With the training parameters shown in Appendix A.1, the best NE performance was obtained from the first annotation approach, where NE beginnings were not explicitly annotated. The lower performance of the second annotation approach can be attributed to the low quantity of training data for the beginning marker annotations. While word level annotation was explored, it is likely to need a far greater quantity of data to enable mapping of different length inputs to a single label."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.59,
+                0.886,
+                0.798
+            ],
+            "angle": 0,
+            "content": "Separately, HuBERT was also fine-tuned for automatic speech recognition (ASR), i.e. for transcribing text from audio. Identical training data was used, with annotation being the transcription provided as part of the NE annotation (with NE annotation removed). The same parameters were employed for its training. Alongside the predicted (NE or ASR) annotation, prediction output also yields an offset which can be converted to a time offset. This can be used to identify the position of the NE(s) to be replaced, and after a greedy alignment of the two outputs, the original transcription of the original NE(s) can be extracted."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.815,
+                0.762,
+                0.831
+            ],
+            "angle": 0,
+            "content": "2.2 Finding an alternative NE"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Once a person NE is identified, a suitable equivalent substitution needs to be obtained, i.e. we want to find the word which could replace the NE in the text if the NE was hidden. This is precisely the concept behind masked language models (MLMs):"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.881,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>The code for the obfuscation pipeline is available from: https://github.com/juditapreiss/speech_ne_obfuscation"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "616"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.115,
+                0.081,
+                0.538,
+                0.374
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.609,
+                0.085,
+                0.859,
+                0.101
+            ],
+            "angle": 0,
+            "content": "Speech to obfuscate as a wav file"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.61,
+                0.152,
+                0.886,
+                0.169
+            ],
+            "angle": 0,
+            "content": "Transcribed text with NEs & timings"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.607,
+                0.221,
+                0.729,
+                0.236
+            ],
+            "angle": 0,
+            "content": "Obfuscated text"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.607,
+                0.254,
+                0.861,
+                0.27
+            ],
+            "angle": 0,
+            "content": "Generated text using cloned voice"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.607,
+                0.287,
+                0.882,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Extracted timings of substituted NEs"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.607,
+                0.321,
+                0.882,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Inserted generated NEs into orig text"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.607,
+                0.356,
+                0.853,
+                0.371
+            ],
+            "angle": 0,
+            "content": "WER and CER based confidence"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.392,
+                0.383,
+                0.604,
+                0.398
+            ],
+            "angle": 0,
+            "content": "Figure 1: Obfuscation pipeline"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.424,
+                0.49,
+                0.68
+            ],
+            "angle": 0,
+            "content": "these models learn their weights so that given a sentence with a hidden (masked) word, the model will output the complete original sentence. The (ASR extracted) original sentences with NEs (as identified by the NE tuned model) masked were passed to a MLM. Three MLM models were explored: BERT, bert-large-uncased model (Devlin et al., 2019), ALBERT, albert-xxlarge-v2, model (Lan et al., 2019) and the distilled RoBERTa base, distilroberta-base, model (Sanh et al., 2019). Each model, with no additional tuning, results in a (pre-specified) number of predictions for each NE in the sentence. Since the models used different datasets in training, their predictions are expected to be different: for example, some may suggest pronouns rather than names."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.683,
+                0.49,
+                0.876
+            ],
+            "angle": 0,
+            "content": "Given the propensity of the MLM to return substitutions which are not names (for example, for the sentence you should call Stella, the model returns you should call him, you should call them, you should call 911 etc), an external list of people names is used for the validation of the proposed suggestions<sup>2</sup> and the highest scoring substitution is returned. Heuristically, the original name is matched against the list to identify whether it is a first or a last name (where possible) and names of the same type suggested by the MLM are returned. Simple rules are employed (last of a sequence of"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.424,
+                0.885,
+                0.6
+            ],
+            "angle": 0,
+            "content": "names is a last name, a single name without a title is a first name etc) to decide on a substitution when the original name does not appear in either the first or last name list. Given the nature of MLMs, suggested alternatives are likely to be more common words: as a positive side effect, this should make them easier to render with voice cloning as they may already appear in the reference speech. Should MLM fail to propose any suitable substitutions, one is selected at random from the first & last name lists, subject to the same heuristic rules."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.62,
+                0.81,
+                0.636
+            ],
+            "angle": 0,
+            "content": "2.3 Reconstruction of original audio"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.646,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this work, the substitute NE is to be re-inserted into the original audio. To reduce the risk of de-identification via the extraction of entities which failed to be identified and therefore stayed in their original form, the substitute entity needs to be produced in the speaker's voice. The YourTTS (Casanova et al., 2021) model, which offers the ability for fine-tuning with less than one minute of speech while achieving good results with reasonable quality, can be used to generate the substitute sentence with all available speech of the speaker provided as reference. Note that it is not necessary to remove the original sentence from the reference data: in fact, its presence may result in more accurate rendering of the substitute sentence. The pre-trained model used in this work (tts_models/multilingual/multi-dataset/your_tts)"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.892,
+                0.489,
+                0.918
+            ],
+            "angle": 0,
+            "content": "2In this work, https://github.com/dominictarr/random-name are used."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.485,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "617"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.212
+            ],
+            "angle": 0,
+            "content": "was trained on the the voice cloning toolkit (VCTK) dataset (Yamagishi et al., 2019) which contains approximately 400 sentence, selected from newspaper text, uttered by 108-110 different speakers, giving it its generalization power. Aside from the reference passed to the model on the command line, no tuning or training of the YourTTS model is done in this work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.214,
+                0.49,
+                0.454
+            ],
+            "angle": 0,
+            "content": "The ASR transcribed text with the substituted NE is generated, rather than the substitution alone, to ensure that the intonation as closely matches the substitution's position in the sentence. The average amplitude of the generated audio is matched to that of the original segment using the Python pydub library. The generated audio is again passed through the HuBERT based NE recognizer, to identify the location of the substituted NE in the generated audio and allow its extraction (note that in this pass, it is not necessary to perform ASR - only the offsets of the replacement NE are required). Should the NE recognizer not identify the same number of NEs as were present in the original, the instance is flagged for manual review."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.456,
+                0.49,
+                0.585
+            ],
+            "angle": 0,
+            "content": "For each NE in the text, a pair of start and end offsets are available: one pair extracted by the HubERT based NE extraction from the original audio and a second pair from the audio generated from the substituted text. This allows the new NEs to be inserted in place of the original NEs. The splicing and concatenation of the waveforms is also performed using the pydub library."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.586,
+                0.49,
+                0.745
+            ],
+            "angle": 0,
+            "content": "A second HuBERT based ASR pass over the newly constructed (substituted) audio, and its comparison against the substituted text using word error rate (WER) and character error rate (CER) gives measures of confidence. Both the metrics, commonly used for evaluation of ASR, allow for sequences of different length to the target – the further the reconstructed audio is from the target sentence, the less likely it is that the substitution will go unnoticed."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.759,
+                0.342,
+                0.773
+            ],
+            "angle": 0,
+            "content": "3 Results and discussion"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.786,
+                0.405,
+                0.8
+            ],
+            "angle": 0,
+            "content": "3.1 Identification of named entities"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The 70,769 training corpus, sampled at \\(16\\mathrm{kHz}\\), is divided up into \\(70\\%\\) for training (49,540 instances), and \\(15\\%\\) for both validation and evaluation (10,615 examples). The hubert-base-1s960 model is used with parameters listed in Appendix A.1. The performance in training, indicated via WER and CER, is shown in Table 1 for both ASR and NER."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.51,
+                0.083,
+                0.765,
+                0.131
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Eval WER</td><td>Eval CER</td></tr><tr><td>ASR</td><td>0.142</td><td>-</td></tr><tr><td>NE</td><td>0.199</td><td>0.063</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.143,
+                0.883,
+                0.171
+            ],
+            "angle": 0,
+            "content": "Table 1: Metric results of the ASR and NE HuBERT based models"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.187,
+                0.895,
+                0.255
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>MLM</td><td>Avg ASR</td><td>NE</td><td>Avg confidence</td></tr><tr><td>ALBERT</td><td>0.980</td><td>13/20</td><td>0.109</td></tr><tr><td>BERT</td><td>0.980</td><td>13/20</td><td>0.098</td></tr><tr><td>RoBERTa</td><td>0.980</td><td>13/20</td><td>0.106</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.563,
+                0.264,
+                0.828,
+                0.278
+            ],
+            "angle": 0,
+            "content": "Table 2: Evaluation of individual steps"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.305,
+                0.884,
+                0.401
+            ],
+            "angle": 0,
+            "content": "For the purpose of the demonstrating the viability of the prototype, no hyperparameter optimization was performed, and the larger HuBERT models were not employed, however improvement in performance of both models are expected should this be pursued."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.414,
+                0.761,
+                0.429
+            ],
+            "angle": 0,
+            "content": "3.2 Finding an alternative NE"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.436,
+                0.885,
+                0.787
+            ],
+            "angle": 0,
+            "content": "A small scale evaluation is performed on a sample of 20 sentences selected at random from the LibriSpeech corpus (Panayotov et al., 2015) across 6 speakers. Sentence selection was subject to them containing a person named entity. While detailed results for the individual steps can be found in Table 2, it should be noted that – for the purposes of this work – the focus is the accuracy of the extraction of the correct NE. The stated accuracy is therefore somewhat misleading: in a number of cases, such as the word Raphael, the named entity is divided into two separate words, suggesting two consecutive named entities. However, this issue is corrected when the NE output is aligned with ASR output and the two separate NE instances are (correctly) merged. Cases with NEs which cannot be aligned are flagged up for manual intervention. The average ASR and (exact match) NE identification do not vary when a different MLM is employed, as this only effects the selection of the substituted name, resulting in different average confidence values."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.802,
+                0.81,
+                0.817
+            ],
+            "angle": 0,
+            "content": "3.3 Reconstruction of original audio"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The voice cloning model requires some reference audio for the speaker: for the 6 selected speakers, 4 have less than 5 audio files (two having 3, and one having only 2 files) in the dataset. The quantity of data used as reference is likely to impact the quality (in terms of its similarity to the original speaker)"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "618"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.232
+            ],
+            "angle": 0,
+            "content": "of the generated text. Given the likely scenarios of deployment, such as dialogues where more than 2 sentences of speech per speaker are available, this may not be representative of the results obtainable with the pipeline. However, it should be noted that even if all substituted instances can be identified as substitutions, the system is equal to a masking technique (where an entity is replaced with a fixed entity, such as a bleep)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.243,
+                0.248,
+                0.258
+            ],
+            "angle": 0,
+            "content": "4 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.269,
+                0.49,
+                0.429
+            ],
+            "angle": 0,
+            "content": "The prototype described shows the steps of an obfuscation pipeline for speech, which results in substituted person named entities uttered in the original speakers voice and replaced in the original audio signal. The prototype makes use of a named entity recognizer built directly on top of audio input, and employs masked language models to generate the substituted entity. It offers an end-to-end automatic solution enabling the sharing of speech with identifying information removed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.43,
+                0.49,
+                0.577
+            ],
+            "angle": 0,
+            "content": "The resulting obfuscated speech remains in the original speaker's voice, allowing for the application of traditional speaker anonymization approaches to mask the speaker's identity. The original prosody can be protected by applying a transformation such as waveform change, offering a significant advantage over a technique which generates a complete obfuscated transcription (instead of splicing an obfuscated entity into original speech)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.587,
+                0.221,
+                0.603
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.613,
+                0.49,
+                0.758
+            ],
+            "angle": 0,
+            "content": "The cloning model used, YourTTS, is trained on the VCTK dataset which consists of high-quality speech signal. It is therefore unclear whether the same accuracy would be obtained with lower quality signal which may contain some background noise. (However, it should again be noted that even if all substituted instances are identifiable in the output, the system is equivalent to a masking model.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.49,
+                0.854
+            ],
+            "angle": 0,
+            "content": "The selection of a person NE replacement does not currently account for continuity: if the same person entity is referred to later, it may be substituted with a different entity to the previous occasion. In addition, the back-off strategy ignores aspects such as gender."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.856,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "To show the approach feasible, very little optimization was performed. Further training and parameter optimization is likely to lead to improved performance for both ASR and NER models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.182
+            ],
+            "angle": 0,
+            "content": "The approach is currently only implemented for person NEs but it could be extended very simply to other types of NEs. However, the degree to which other entity types require obfuscation in speech is not clear to us as mentions of organizations may well not be identifying at all."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.198,
+                0.662,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Ethics Statement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.226,
+                0.885,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Aside from the ethical concerns regarding voice cloning (covered in e.g. YourTTS (Casanova et al., 2021)), deployment would require a detailed evaluation of risk of de-identification. It is believed that the final confidence and the accuracy of each step can be combined to significantly reduce this risk. The voice itself also offers options for identification: the value of yielding substitutions in the original speaker's voice (and keeping the original prosody) would need to be weighed up against approaches which anonymize voice but preserve prosodic information."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.45,
+                0.61,
+                0.465
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.475,
+                0.884,
+                0.529
+            ],
+            "angle": 0,
+            "content": "Mohamed Abdalla, Moustafa Abdalla, Frank Rudzicz, and Graeme Hirsto. 2020. Using word embeddings to improve the privacy of clinical notes. J Am Med Inform Assoc, 27(6):901-907."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.543,
+                0.885,
+                0.61
+            ],
+            "angle": 0,
+            "content": "Edresson Casanova, Julian Weber, Christopher Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir Antonelli Ponti. 2021. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. CoRR, abs/2112.02418."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.624,
+                0.884,
+                0.678
+            ],
+            "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-HLT."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.691,
+                0.885,
+                0.719
+            ],
+            "angle": 0,
+            "content": "Gölge Eren and The Coqui TTS Team. 2021. Coqui TTS. https://github.com/coqui-ai/TTS."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.733,
+                0.885,
+                0.839
+            ],
+            "angle": 0,
+            "content": "Tzvika Hartman, Michael D. Howell, Jeff Dean, Shlomo Hoory, Ronit Slyper, Itay Laish, Oren Gilon, Danny Vainstein, Greg Corrado, Katherine Chou, Ming Jack Po, Jutta Williams, Scott Ellis, Gavin Bee, Avinatan Hassidim, Rony Amira, Genady Beryozkin, Idan Szpektor, and Yossi Matias. 2020. Customization scenarios for de-identification of clinical notes. BMC Med Inform Decis Mak, 20(14)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.852,
+                0.887,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. CoRR, abs/2106.07447."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.475,
+                0.887,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "619"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.086,
+                0.49,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. ALBERT: A lite BERT for self-supervised learning of language representations. CoRR, abs/1909.11942."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.161,
+                0.49,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Pierre Lison, Ildiko Pilán, David Sanchez, Montserrat Batet, and Lilja Øvrelid. 2021. Anonymisation models for text data: State of the art, challenges and future directions. 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 4188-4203, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.288,
+                0.49,
+                0.46
+            ],
+            "angle": 0,
+            "content": "Arttu Oksanen, Eero Hyvönen, Minna Tamper, Jouni Tuominen, Henna Ylimaa, Katja Löytynoja, Matti Kokkonen, and Aki Hietanen. 2022. An anonymization tool for open data publication of legal documents. In Joint Proceedings of the 3th International Workshop on Artificial Intelligence Technologies for Legal Documents (AI4LEGAL 2022) and the 1st International Workshop on Knowledge Graph Summarization (KGSum 2022) co-located with the 21st International Semantic Web Conference (ISWC 2022), Virtual Event, Hangzhou, China, October 23-24, 2022, volume 3257 of CEUR Workshop Proceedings, pages 12-21. CEUR-WS.org."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.467,
+                0.49,
+                0.534
+            ],
+            "angle": 0,
+            "content": "Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an ASR corpus based on public domain audio books. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.542,
+                0.49,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.605,
+                0.49,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Samuel Sousa and Roman Kern. 2022. How to keep text private? a systematic review of deep learning methods for privacy-preserving natural language processing. Artif Intell Rev."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.666,
+                0.49,
+                0.745
+            ],
+            "angle": 0,
+            "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.754,
+                0.489,
+                0.796
+            ],
+            "angle": 0,
+            "content": "Hemant Yadav, Sreyan Ghosh, Yi Yu, and Rajiv Ratn Shah. 2020. End-to-end named entity recognition from English speech. CoRR, abs/2005.11184."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.804,
+                0.49,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Junichi Yamagishi, Christophe Veaux, and Kirsten MacDonald. 2019. CSTR VCTK corpus: English multispeaker corpus for CSTR voice cloning toolkit (version 0.92)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.866,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Haitong Zhang and Yue Lin. 2022. Improve few-shot voice cloning using multi-modal learning. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 8317-8321."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.115,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.742,
+                0.101
+            ],
+            "angle": 0,
+            "content": "A Model training details"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.11,
+                0.73,
+                0.126
+            ],
+            "angle": 0,
+            "content": "A.1 HuBERT parameters"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.131,
+                0.885,
+                0.195
+            ],
+            "angle": 0,
+            "content": "HubertForCTC hubert-base-1s960 model was used with frozen feature encoder, training size of 49540 examples, and validation and evaluation size of 10615, alongside the following parameters:"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.53,
+                0.197,
+                0.877,
+                0.424
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Parameter</td><td>Value</td></tr><tr><td>group_by_length</td><td>True</td></tr><tr><td>per_device_train_batch_size</td><td>8</td></tr><tr><td>per_device_eval_batch_size</td><td>batch_size</td></tr><tr><td>evaluation_strategy</td><td>&quot;steps&quot;</td></tr><tr><td>num_train_epochs</td><td>num_epochs</td></tr><tr><td>fp16</td><td>True</td></tr><tr><td>gradient_checkpointing</td><td>True</td></tr><tr><td>save_steps</td><td>500</td></tr><tr><td>eval_steps</td><td>500</td></tr><tr><td>learning_rate</td><td>1e-4</td></tr><tr><td>weight Decay</td><td>0.005</td></tr><tr><td>warmup_steps</td><td>1000</td></tr><tr><td>num_epochs</td><td>30</td></tr></table>"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "620"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.133,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.127,
+                0.533,
+                0.144
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.145,
+                0.397,
+                0.159
+            ],
+            "angle": 0,
+            "content": "Unnumbered, follows conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.17,
+                0.553,
+                0.187
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.188,
+                0.753,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Ethics Statement discusses potential risks of the work. This follows \"Limitations\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.213,
+                0.697,
+                0.229
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.231,
+                0.223,
+                0.244
+            ],
+            "angle": 0,
+            "content": "Section 1"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.257,
+                0.67,
+                0.273
+            ],
+            "angle": 0,
+            "content": "A4. Have you used AI writing assistants when working on this paper?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.274,
+                0.233,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.301,
+                0.49,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.322,
+                0.215,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.348,
+                0.53,
+                0.364
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.365,
+                0.249,
+                0.38
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.391,
+                0.779,
+                0.407
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.408,
+                0.249,
+                0.423
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.434,
+                0.882,
+                0.497
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.499,
+                0.249,
+                0.514
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.525,
+                0.882,
+                0.573
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.575,
+                0.249,
+                0.589
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.6,
+                0.882,
+                0.632
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.634,
+                0.249,
+                0.648
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.659,
+                0.882,
+                0.74
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.741,
+                0.249,
+                0.755
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.765,
+                0.495,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.787,
+                0.246,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Sections 2 & 3"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.813,
+                0.882,
+                0.845
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.847,
+                0.882,
+                0.877
+            ],
+            "angle": 0,
+            "content": "Appendix A1. The infrastructure was chosen by the HPC and therefore wasn't known - if needed, the experiments can be rerun with this constrained."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.885,
+                0.878,
+                0.909
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.514,
+                0.941
+            ],
+            "angle": 0,
+            "content": "621"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.084,
+                0.88,
+                0.116
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.148,
+                0.118,
+                0.882,
+                0.149
+            ],
+            "angle": 0,
+            "content": "Experimental setup discussed, hyperparameter optimization not performed as stated as the paper only checks viability."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.159,
+                0.882,
+                0.207
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.148,
+                0.209,
+                0.456,
+                0.224
+            ],
+            "angle": 0,
+            "content": "Average used, clearly denoted. Section 3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.235,
+                0.882,
+                0.282
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.284,
+                0.351,
+                0.299
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.31,
+                0.878,
+                0.327
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.331,
+                0.216,
+                0.347
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.358,
+                0.882,
+                0.39
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.391,
+                0.25,
+                0.406
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.417,
+                0.882,
+                0.464
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.466,
+                0.25,
+                0.481
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.492,
+                0.882,
+                0.539
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.541,
+                0.25,
+                0.556
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.567,
+                0.874,
+                0.582
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.584,
+                0.25,
+                0.599
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.61,
+                0.882,
+                0.64
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.642,
+                0.25,
+                0.657
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "622"
+        }
+    ]
+]

2023/Automatic Named Entity Obfuscation in Speech/f423cca1-1aeb-43bd-8d25-5845a22e4c4c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e92712c394761baa8f0a67e53772d134c69282f9eaf3c9e01755bac6f7ac0d2
+size 155223

2023/Automatic Named Entity Obfuscation in Speech/full.md

@@ -0,0 +1,237 @@
+# Automatic Named Entity Obfuscation in Speech
+
+Judita Preiss  
+University of Sheffield, Information School  
+The Wave, 2 Whitham Road  
+Sheffield S10 2AH  
+judita.preiss@sheffield.ac.uk
+
+# Abstract
+
+Sharing data containing personal information often requires its anonymization, even when consent for sharing was obtained from the data originator. While approaches exist for automated anonymization of text, the area is not as thoroughly explored in speech. This work focuses on identifying, replacing and inserting replacement named entities synthesized using voice cloning into original audio thereby retaining prosodic information while reducing the likelihood of deanonymization. The approach employs a novel named entity recognition (NER) system built directly on speech by training HuBERT (Hsu et al., 2021) using the English speech NER dataset (Yadav et al., 2020). Name substitutes are found using a masked language model and are synthesized using text to speech voice cloning (Eren and Team, 2021), upon which the substitute named entities are re-inserted into the original text. The approach is prototyped on a sample of the LibriSpeech corpus (Panayotov et al., 2015) with each step evaluated individually.
+
+# 1 Introduction
+
+Privacy concerns, particularly where an individual could be identified, preclude sharing and therefore automatic exploitation of many data sources. Anonymization, the removal of identifying information, has been automated for text (Lison et al., 2021), including large scale applications such as in clinical (Hartman et al., 2020) or legal settings (Oksanen et al., 2022), with off-the-shelf systems having reported performance of  $90 + \%$  (Hartman et al., 2020). To minimize the risk of re-identification, obfuscation - replacing identifying information with a different substitute of the same type - has been explored as an alternative to replacing identifying information with a generic marker (Sousa and Kern, 2022). The main focus in speech has been on voice anonymization, which may not be a problem with speaker consent, with the removal of
+
+identifying information receiving less attention. To our knowledge, this is the first prototype to perform named entity obfuscation directly, in the original speaker's voice. Aside from voice cloning, it explores a named entity recognition approach based directly on audio signal and uses language model masking to find appropriate substitutions.
+
+Recent advances in speech models, particularly the inclusion of language models within the speech model itself (e.g. HuBERT (Hsu et al., 2021)) gives models greater insight into expected contexts. Previous work on named entity recognition (NER) in speech frequently employs a two step approach, transcribing speech first, followed by the application of existing named entity techniques (Yadav et al., 2020). However, this process has the potential to compound errors as errors in transcription will increase the probability of error in NER. We suggest that the addition of language models into the speech model gives these sufficient power to perform NER directly, and therefore that transcribing (automatic speech recognition, ASR) and NER can be separated, and used to provide a confidence measure in their performance. Divided, the two do not propagate errors in the same way; in fact, treating ASR and NER separately allows one to fix (some of the) errors of the other. The proposed second (final) ASR pass merely produces a confidence value in the result to decide whether a manual check should be performed.
+
+The success of few shot learning, where a limited number of examples is used to generalize a pre-trained deep learning model to a new situation, for text-to-speech – and specifically voice cloning (Zhang and Lin, 2022) – enables an alternative, equivalent but different, entity to be inserted in the audio signal in place of the original while preserving the prosody information throughout. While large databases of potential replacement entities can be used to select a substitution, these may not preserve necessary properties (such as gender). Al-
+
+ternatively, word embeddings have been used to suggest close (in the multi-dimensional space) alternatives (Abdalla et al., 2020), however these can suffer from the same drawback. We propose using a more contextualized alternative to word embeddings, a masked language model (Devlin et al., 2019), where the model is trained by hiding (masking) words and predictions of the original word are made based on their context.
+
+This work makes the following contributions: (1) a complete obfuscation pipeline for names in speech $^{1}$ , (2) a named entity recognizer built directly on speech without requiring text transcription first, (3) alternative (obfuscated) entity replacement selection via masking language model, and (4) confidence annotated system output, allowing for manual correction and / or selection of shareable instances. Section 2 contains the methodology with results in Section 3. Section 4 presents the conclusions and future work.
+
+# 2 Methodology
+
+The steps of the overall pipeline, which takes in an audio file and produces an obfuscated audio file along with a confidence value, can be found in Figure 1. The approach comprises of three main parts: 1) identification of named entities (NEs) in the audio, 2) finding an equivalent alternative for the original NEs, and 3) reconstructing the original audio to incorporate the replacement NEs. The reconstructed audio can further be used to obtain a confidence value.
+
+# 2.1 Identification of named entities
+
+To enable the direct use of a language model on speech input for the purpose of named entity recognition (NER), a dataset of audio recordings with annotated NEs is required. The English speech NER dataset (Yadav et al., 2020), which consists of 70,769 waveforms with transcripts annotated with person, location and organization NEs, is used for fine-tuning the Hidden-Unit BERT speech model (HuBERT) (Hsu et al., 2021). HuBERT was selected over other speech models since it learns both acoustic and language models from its inputs and therefore has an increased awareness of context. The success of language models on text NER has demonstrated how crucial context is for this
+
+task, and using a model which incorporates both an acoustic and a language model (over acoustic only) allows the approach to exploit the information used in text NER, while managing to avoid the need for a transcript.
+
+For training, NE annotations need to be converted to a suitable format, indicating the presence or absence of a NE in each position. Following the inside-outside(-beginning) chunking common to many NER approaches (Tjong Kim Sang and De Meulder, 2003), three formats were explored: 1) character level annotation, mapping each character to either  $o$  for a character outside of a named entity, space, or  $n$ ,  $l$ ,  $e$  for characters within person, location or organization entities respectively, 2) the same character level annotation with separate characters added to denote the beginning of each type of NE (mapping the sentence TELL JACK to oooo mnnn with  $m$  denoting the start of a person NE), 3) and, for completeness, annotation was also explored at word level.
+
+With the training parameters shown in Appendix A.1, the best NE performance was obtained from the first annotation approach, where NE beginnings were not explicitly annotated. The lower performance of the second annotation approach can be attributed to the low quantity of training data for the beginning marker annotations. While word level annotation was explored, it is likely to need a far greater quantity of data to enable mapping of different length inputs to a single label.
+
+Separately, HuBERT was also fine-tuned for automatic speech recognition (ASR), i.e. for transcribing text from audio. Identical training data was used, with annotation being the transcription provided as part of the NE annotation (with NE annotation removed). The same parameters were employed for its training. Alongside the predicted (NE or ASR) annotation, prediction output also yields an offset which can be converted to a time offset. This can be used to identify the position of the NE(s) to be replaced, and after a greedy alignment of the two outputs, the original transcription of the original NE(s) can be extracted.
+
+# 2.2 Finding an alternative NE
+
+Once a person NE is identified, a suitable equivalent substitution needs to be obtained, i.e. we want to find the word which could replace the NE in the text if the NE was hidden. This is precisely the concept behind masked language models (MLMs):
+
+Figure 1: Obfuscation pipeline  
+![](images/486fcf9d398058710632011d88c10eff46b67d5226d5edf62beb14bd3d9045d1.jpg)  
+Speech to obfuscate as a wav file  
+Transcribed text with NEs & timings  
+Obfuscated text  
+Generated text using cloned voice  
+Extracted timings of substituted NEs  
+Inserted generated NEs into orig text  
+WER and CER based confidence
+
+these models learn their weights so that given a sentence with a hidden (masked) word, the model will output the complete original sentence. The (ASR extracted) original sentences with NEs (as identified by the NE tuned model) masked were passed to a MLM. Three MLM models were explored: BERT, bert-large-uncased model (Devlin et al., 2019), ALBERT, albert-xxlarge-v2, model (Lan et al., 2019) and the distilled RoBERTa base, distilroberta-base, model (Sanh et al., 2019). Each model, with no additional tuning, results in a (pre-specified) number of predictions for each NE in the sentence. Since the models used different datasets in training, their predictions are expected to be different: for example, some may suggest pronouns rather than names.
+
+Given the propensity of the MLM to return substitutions which are not names (for example, for the sentence you should call Stella, the model returns you should call him, you should call them, you should call 911 etc), an external list of people names is used for the validation of the proposed suggestions<sup>2</sup> and the highest scoring substitution is returned. Heuristically, the original name is matched against the list to identify whether it is a first or a last name (where possible) and names of the same type suggested by the MLM are returned. Simple rules are employed (last of a sequence of
+
+names is a last name, a single name without a title is a first name etc) to decide on a substitution when the original name does not appear in either the first or last name list. Given the nature of MLMs, suggested alternatives are likely to be more common words: as a positive side effect, this should make them easier to render with voice cloning as they may already appear in the reference speech. Should MLM fail to propose any suitable substitutions, one is selected at random from the first & last name lists, subject to the same heuristic rules.
+
+# 2.3 Reconstruction of original audio
+
+In this work, the substitute NE is to be re-inserted into the original audio. To reduce the risk of de-identification via the extraction of entities which failed to be identified and therefore stayed in their original form, the substitute entity needs to be produced in the speaker's voice. The YourTTS (Casanova et al., 2021) model, which offers the ability for fine-tuning with less than one minute of speech while achieving good results with reasonable quality, can be used to generate the substitute sentence with all available speech of the speaker provided as reference. Note that it is not necessary to remove the original sentence from the reference data: in fact, its presence may result in more accurate rendering of the substitute sentence. The pre-trained model used in this work (tts_models/multilingual/multi-dataset/your_tts)
+
+was trained on the the voice cloning toolkit (VCTK) dataset (Yamagishi et al., 2019) which contains approximately 400 sentence, selected from newspaper text, uttered by 108-110 different speakers, giving it its generalization power. Aside from the reference passed to the model on the command line, no tuning or training of the YourTTS model is done in this work.
+
+The ASR transcribed text with the substituted NE is generated, rather than the substitution alone, to ensure that the intonation as closely matches the substitution's position in the sentence. The average amplitude of the generated audio is matched to that of the original segment using the Python pydub library. The generated audio is again passed through the HuBERT based NE recognizer, to identify the location of the substituted NE in the generated audio and allow its extraction (note that in this pass, it is not necessary to perform ASR - only the offsets of the replacement NE are required). Should the NE recognizer not identify the same number of NEs as were present in the original, the instance is flagged for manual review.
+
+For each NE in the text, a pair of start and end offsets are available: one pair extracted by the HubERT based NE extraction from the original audio and a second pair from the audio generated from the substituted text. This allows the new NEs to be inserted in place of the original NEs. The splicing and concatenation of the waveforms is also performed using the pydub library.
+
+A second HuBERT based ASR pass over the newly constructed (substituted) audio, and its comparison against the substituted text using word error rate (WER) and character error rate (CER) gives measures of confidence. Both the metrics, commonly used for evaluation of ASR, allow for sequences of different length to the target – the further the reconstructed audio is from the target sentence, the less likely it is that the substitution will go unnoticed.
+
+# 3 Results and discussion
+
+# 3.1 Identification of named entities
+
+The 70,769 training corpus, sampled at  $16\mathrm{kHz}$ , is divided up into  $70\%$  for training (49,540 instances), and  $15\%$  for both validation and evaluation (10,615 examples). The hubert-base-1s960 model is used with parameters listed in Appendix A.1. The performance in training, indicated via WER and CER, is shown in Table 1 for both ASR and NER.
+
+<table><tr><td></td><td>Eval WER</td><td>Eval CER</td></tr><tr><td>ASR</td><td>0.142</td><td>-</td></tr><tr><td>NE</td><td>0.199</td><td>0.063</td></tr></table>
+
+Table 1: Metric results of the ASR and NE HuBERT based models  
+
+<table><tr><td>MLM</td><td>Avg ASR</td><td>NE</td><td>Avg confidence</td></tr><tr><td>ALBERT</td><td>0.980</td><td>13/20</td><td>0.109</td></tr><tr><td>BERT</td><td>0.980</td><td>13/20</td><td>0.098</td></tr><tr><td>RoBERTa</td><td>0.980</td><td>13/20</td><td>0.106</td></tr></table>
+
+Table 2: Evaluation of individual steps
+
+For the purpose of the demonstrating the viability of the prototype, no hyperparameter optimization was performed, and the larger HuBERT models were not employed, however improvement in performance of both models are expected should this be pursued.
+
+# 3.2 Finding an alternative NE
+
+A small scale evaluation is performed on a sample of 20 sentences selected at random from the LibriSpeech corpus (Panayotov et al., 2015) across 6 speakers. Sentence selection was subject to them containing a person named entity. While detailed results for the individual steps can be found in Table 2, it should be noted that – for the purposes of this work – the focus is the accuracy of the extraction of the correct NE. The stated accuracy is therefore somewhat misleading: in a number of cases, such as the word Raphael, the named entity is divided into two separate words, suggesting two consecutive named entities. However, this issue is corrected when the NE output is aligned with ASR output and the two separate NE instances are (correctly) merged. Cases with NEs which cannot be aligned are flagged up for manual intervention. The average ASR and (exact match) NE identification do not vary when a different MLM is employed, as this only effects the selection of the substituted name, resulting in different average confidence values.
+
+# 3.3 Reconstruction of original audio
+
+The voice cloning model requires some reference audio for the speaker: for the 6 selected speakers, 4 have less than 5 audio files (two having 3, and one having only 2 files) in the dataset. The quantity of data used as reference is likely to impact the quality (in terms of its similarity to the original speaker)
+
+of the generated text. Given the likely scenarios of deployment, such as dialogues where more than 2 sentences of speech per speaker are available, this may not be representative of the results obtainable with the pipeline. However, it should be noted that even if all substituted instances can be identified as substitutions, the system is equal to a masking technique (where an entity is replaced with a fixed entity, such as a bleep).
+
+# 4 Conclusion
+
+The prototype described shows the steps of an obfuscation pipeline for speech, which results in substituted person named entities uttered in the original speakers voice and replaced in the original audio signal. The prototype makes use of a named entity recognizer built directly on top of audio input, and employs masked language models to generate the substituted entity. It offers an end-to-end automatic solution enabling the sharing of speech with identifying information removed.
+
+The resulting obfuscated speech remains in the original speaker's voice, allowing for the application of traditional speaker anonymization approaches to mask the speaker's identity. The original prosody can be protected by applying a transformation such as waveform change, offering a significant advantage over a technique which generates a complete obfuscated transcription (instead of splicing an obfuscated entity into original speech).
+
+# Limitations
+
+The cloning model used, YourTTS, is trained on the VCTK dataset which consists of high-quality speech signal. It is therefore unclear whether the same accuracy would be obtained with lower quality signal which may contain some background noise. (However, it should again be noted that even if all substituted instances are identifiable in the output, the system is equivalent to a masking model.)
+
+The selection of a person NE replacement does not currently account for continuity: if the same person entity is referred to later, it may be substituted with a different entity to the previous occasion. In addition, the back-off strategy ignores aspects such as gender.
+
+To show the approach feasible, very little optimization was performed. Further training and parameter optimization is likely to lead to improved performance for both ASR and NER models.
+
+The approach is currently only implemented for person NEs but it could be extended very simply to other types of NEs. However, the degree to which other entity types require obfuscation in speech is not clear to us as mentions of organizations may well not be identifying at all.
+
+# Ethics Statement
+
+Aside from the ethical concerns regarding voice cloning (covered in e.g. YourTTS (Casanova et al., 2021)), deployment would require a detailed evaluation of risk of de-identification. It is believed that the final confidence and the accuracy of each step can be combined to significantly reduce this risk. The voice itself also offers options for identification: the value of yielding substitutions in the original speaker's voice (and keeping the original prosody) would need to be weighed up against approaches which anonymize voice but preserve prosodic information.
+
+# References
+
+Mohamed Abdalla, Moustafa Abdalla, Frank Rudzicz, and Graeme Hirsto. 2020. Using word embeddings to improve the privacy of clinical notes. J Am Med Inform Assoc, 27(6):901-907.  
+Edresson Casanova, Julian Weber, Christopher Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir Antonelli Ponti. 2021. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. CoRR, abs/2112.02418.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT.  
+Gölge Eren and The Coqui TTS Team. 2021. Coqui TTS. https://github.com/coqui-ai/TTS.  
+Tzvika Hartman, Michael D. Howell, Jeff Dean, Shlomo Hoory, Ronit Slyper, Itay Laish, Oren Gilon, Danny Vainstein, Greg Corrado, Katherine Chou, Ming Jack Po, Jutta Williams, Scott Ellis, Gavin Bee, Avinatan Hassidim, Rony Amira, Genady Beryozkin, Idan Szpektor, and Yossi Matias. 2020. Customization scenarios for de-identification of clinical notes. BMC Med Inform Decis Mak, 20(14).  
+Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. CoRR, abs/2106.07447.
+
+Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. ALBERT: A lite BERT for self-supervised learning of language representations. CoRR, abs/1909.11942.  
+Pierre Lison, Ildiko Pilán, David Sanchez, Montserrat Batet, and Lilja Øvrelid. 2021. Anonymisation models for text data: State of the art, challenges and future directions. 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 4188-4203, Online. Association for Computational Linguistics.  
+Arttu Oksanen, Eero Hyvönen, Minna Tamper, Jouni Tuominen, Henna Ylimaa, Katja Löytynoja, Matti Kokkonen, and Aki Hietanen. 2022. An anonymization tool for open data publication of legal documents. In Joint Proceedings of the 3th International Workshop on Artificial Intelligence Technologies for Legal Documents (AI4LEGAL 2022) and the 1st International Workshop on Knowledge Graph Summarization (KGSum 2022) co-located with the 21st International Semantic Web Conference (ISWC 2022), Virtual Event, Hangzhou, China, October 23-24, 2022, volume 3257 of CEUR Workshop Proceedings, pages 12-21. CEUR-WS.org.  
+Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an ASR corpus based on public domain audio books. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE.  
+Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108.  
+Samuel Sousa and Roman Kern. 2022. How to keep text private? a systematic review of deep learning methods for privacy-preserving natural language processing. Artif Intell Rev.  
+Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147.  
+Hemant Yadav, Sreyan Ghosh, Yi Yu, and Rajiv Ratn Shah. 2020. End-to-end named entity recognition from English speech. CoRR, abs/2005.11184.  
+Junichi Yamagishi, Christophe Veaux, and Kirsten MacDonald. 2019. CSTR VCTK corpus: English multispeaker corpus for CSTR voice cloning toolkit (version 0.92).  
+Haitong Zhang and Yue Lin. 2022. Improve few-shot voice cloning using multi-modal learning. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 8317-8321.
+
+# A Model training details
+
+# A.1 HuBERT parameters
+
+HubertForCTC hubert-base-1s960 model was used with frozen feature encoder, training size of 49540 examples, and validation and evaluation size of 10615, alongside the following parameters:
+
+<table><tr><td>Parameter</td><td>Value</td></tr><tr><td>group_by_length</td><td>True</td></tr><tr><td>per_device_train_batch_size</td><td>8</td></tr><tr><td>per_device_eval_batch_size</td><td>batch_size</td></tr><tr><td>evaluation_strategy</td><td>&quot;steps&quot;</td></tr><tr><td>num_train_epochs</td><td>num_epochs</td></tr><tr><td>fp16</td><td>True</td></tr><tr><td>gradient_checkpointing</td><td>True</td></tr><tr><td>save_steps</td><td>500</td></tr><tr><td>eval_steps</td><td>500</td></tr><tr><td>learning_rate</td><td>1e-4</td></tr><tr><td>weight Decay</td><td>0.005</td></tr><tr><td>warmup_steps</td><td>1000</td></tr><tr><td>num_epochs</td><td>30</td></tr></table>
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Unnumbered, follows conclusion
+
+A2. Did you discuss any potential risks of your work?
+
+Ethics Statement discusses potential risks of the work. This follows "Limitations".
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Section 1
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+No response.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+No response.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+No response.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+No response.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+No response.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+No response.
+
+C Did you run computational experiments?
+
+Sections 2 & 3
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Appendix A1. The infrastructure was chosen by the HPC and therefore wasn't known - if needed, the experiments can be rerun with this constrained.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Experimental setup discussed, hyperparameter optimization not performed as stated as the paper only checks viability.
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+Average used, clearly denoted. Section 3.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Not applicable. Left blank.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Automatic Named Entity Obfuscation in Speech/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7265a1dee21a001c91fffd62a0512da87ff68ae08b9637e81da476bdeddf4f28
+size 110764

2023/Automatic Readability Assessment for Closely Related Languages/44092f05-688c-428e-a5e8-3637a992286e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0b2923b471efbf13b4f77e4b9266d373eb2edaf9ffb96e39a3f33e1a4580d4a
+size 488218

2023/Automatic Readability Assessment for Closely Related Languages/full.md

@@ -0,0 +1,348 @@
+# Automatic Readability Assessment for Closely Related Languages
+
+Joseph Marvin Imperial $\Omega, \Lambda$  Ekaterina Kochmar $^{\Upsilon}$ $\Omega$ National University, Philippines  $\Lambda$ University of Bath, UK  
+ $\Upsilon$ MBZUAI, UAE  
+jmri20@bath.ac.uk ekaterina.kochmar@mbzuai.ac.ae
+
+# Abstract
+
+In recent years, the main focus of research on automatic readability assessment (ARA) has shifted towards using expensive deep learning-based methods with the primary goal of increasing models' accuracy. This, however, is rarely applicable for low-resource languages where traditional handcrafted features are still widely used due to the lack of existing NLP tools to extract deeper linguistic representations. In this work, we take a step back from the technical component and focus on how linguistic aspects such as mutual intelligibility or degree of language relatedness can improve ARA in a low-resource setting. We collect short stories written in three languages in the Philippines – Tagalog, Bikol, and Cebuano – to train readability assessment models and explore the interaction of data and features in various cross-lingual setups. Our results show that the inclusion of CROSSNGO, a novel specialized feature exploiting n-gram overlap applied to languages with high mutual intelligibility, significantly improves the performance of ARA models compared to the use of off-the-shelf large multilingual language models alone. Consequently, when both linguistic representations are combined, we achieve state-of-the-art results for Tagalog and Cebuano, and baseline scores for ARA in Bikol.
+
+We release our data and code at github.com/imperialite/ara-close-lang
+
+# 1 Introduction
+
+Automatic readability assessment (ARA) is the task that aims to approximate the difficulty level of a piece of literary material using computer-aided tools. The need for such application arises from challenges related to the misalignment of difficulty labels when humans with various domain expertise provide annotations, as well as to the difficulty of manual extraction of complex text-based features (Deutsch et al., 2020). At the same time,
+
+readability assessment tools often use different definitions of complexity levels based on (a) age level (Vajjala and Meurers, 2012; Xia et al., 2016), (b) grade level (Imperial and Ong, 2020, 2021a), or on established frameworks such as (c) the Common European Framework of Reference for Languages (CEFR) $^{1}$  (François and Fairon, 2012; Pilán et al., 2016; Xia et al., 2016; Reynolds, 2016; Vajjala and Rama, 2018).
+
+In recent years, deep learning methods and large language models (LLMs) have gained popularity in the research community. Often studies using these methodologies focus primarily on improving the performance across various metrics. This is particularly manifest in ARA research in languages with a high number of accessible and publicly-available readability corpora such as English (Heilman et al., 2008; Flor et al., 2013; Vajjala and Lucic, 2018) and German (Hancke et al., 2012; Weiss et al., 2021; Weiss and Meurers, 2022) to name a few. At the same time, existing studies focusing on low-resource languages such as Cebuano (Imperial et al., 2022) and Bengala (Islam et al., 2012; Islam and Rahman, 2014) are still at the stage of primarily using traditional features such as word and sentence lengths to train predictive models.
+
+We identify two problems that are related to the use of complex neural-based approaches: the success of such models depends on (a) whether there is enough available data to train a model using a customized deep neural network, and (b) in the case of LLMs, whether there exists an available off-the-shelf pre-trained model for a low-resource language of interest. Imperial et al. (2022) have recently shown that merely integrating extracted embeddings from a multilingual BERT model as features for Cebuano, a low-resource Philippine language, does not outperform models trained with
+
+orthographic features such as syllable patterns customized for the language. These challenges provide motivation for researchers to further explore methods that do not rely on the availability of large amounts of data or complex pre-trained models and investigate simpler, more interpretable models instead of black box architectures.
+
+In this paper, we take a step back and focus on the data available for low-resource Philippine languages and the features extracted from them rather than on the algorithmic aspects. Specifically, we explore a scenario where small readability corpora are available for languages that are closely related or belong to one major language family tree. To the best of our knowledge, incorporating the degree of language closeness or relatedness has not been explored before in any cross-lingual ARA setup. In this study, we make the following contributions:
+
+1. We conduct an extensive pioneer study on readability assessment in a cross-lingual setting using three closely related Philippine languages: Tagalog, Bikolano, and Cebuano.  
+2. We extract various feature sets ranging from linguistically motivated to neural embeddings, and empirically evaluate how they affect the performance of readability models in a singular, pairwise, and full cross-lingual setup.  
+3. We introduce cross-lingual Character N-gram Overlap (CROSSNGO), a novel feature applicable to readability assessment in closely related languages.  
+4. We also introduce and release a new readability corpus for Bikolano, one of the major languages in the Philippines.  
+5. Finally, we set a baseline for ARA in Bikol and report state-of-the-art results for Tagalog and Cebuano.
+
+# 2 Background
+
+# 2.1 The Philippine Linguistic Profile
+
+The Philippines is a linguistically diverse country in Southeast Asia (SEA) with over 180 languages spoken by over 100 million people. Languages in the Philippines can be best described as morphologically rich due to their free-word order structures and high number of possible inflections, full and partial duplications, and compound words (Go and
+
+Nocon, 2017). In addition, following lexicostatistical studies, languages are divided into two subgroups, northern and central, wherein the major languages Ilokano, Pangasinan, and Kapampangan belong to the northern subgroup, and Tagalog, Bikol, Hiligaynon, and Cebuano are allocated to the central subgroup (Walton, 1979; Constantino, 1998). Figure 1 illustrates the central subgroup of the Philippine language family tree.
+
+![](images/4df3b8c5dd76c9525a350bf76862b13f210998216f1b4493c9b93bb7d4f2bffc.jpg)  
+Figure 1: The central subgroup of the Philippine language family tree highlighting the origins of Tagalog, Bikol, and Cebuano.
+
+In this study, our readability experiments focus on three major Philippine languages, Tagalog, Cebuano, and Bikol, which we refer to further in the paper with their corresponding ISO-639-2 language codes as TGL, CEB, and BCL, respectively.
+
+# 2.2 Mutual Intelligibility
+
+Preliminary linguistic profiling studies of the main Philippine languages such as by McFarland (2004) show that Tagalog, Bikol, and Cebuano are more closely related to one another than any languages in the northern family tree. A language's closeness or its degree of relatedness to another language from the same family (sub)tree is commonly referred to as mutual intelligibility (Bloomfield, 1926). Such similarities can be seen across multiple aspects, including, for example (a) syllable patterns where all three languages have similar three case-marking particles - ang (En: the), ng (En: of), and sa (En: at) for Bikol and Tagalog, and ug instead of sa for Cebuano; and (b) shared words, e.g. mata (En: eye) and tubig (En: water).
+
+For languages belonging to one greater subgroup in the case of Central Philippine for Tagalog, Bikol, and Cebuano, showing stronger quantitative evidence of mutual intelligibility may provide additional proof that these languages are indeed, at some level, closely related to each other. Thus,
+
+to contribute towards further understanding of mutual intelligibility in the Philippines language space, we apply two linguistic similarity-based measures using character n-gram overlap and genetic distance which we discuss in the sections below.
+
+<table><tr><td></td><td>TGL</td><td>BCL</td><td>CEB</td><td>ENG</td></tr><tr><td>TGL</td><td>1.000</td><td>0.810</td><td>0.812</td><td>0.270</td></tr><tr><td>BCL</td><td>0.810</td><td>1.000</td><td>0.789</td><td>0.263</td></tr><tr><td>CEB</td><td>0.812</td><td>0.789</td><td>1.000</td><td>0.213</td></tr><tr><td>ENG</td><td>0.270</td><td>0.263</td><td>0.213</td><td>1.000</td></tr></table>
+
+(a) Bigram Character Overlap  
+
+<table><tr><td></td><td>TGL</td><td>BCL</td><td>CEB</td><td>ENG</td></tr><tr><td>TGL</td><td>1.000</td><td>0.588</td><td>0.628</td><td>0.121</td></tr><tr><td>BCL</td><td>0.588</td><td>1.000</td><td>0.533</td><td>0.144</td></tr><tr><td>CEB</td><td>0.628</td><td>0.533</td><td>1.000</td><td>0.090</td></tr><tr><td>ENG</td><td>0.121</td><td>0.144</td><td>0.090</td><td>1.000</td></tr></table>
+
+Character N-Gram Overlap. For our first measure, we use the overlap in character bigrams and trigrams for every pair from the selected set of languages. To do this, we simply extract and rank the top occurring character bigrams and trigrams for a given language and calculate the Rank-Biased Overlap  $(\mathrm{RBO})^2$  (Webber et al., 2010). RBO provides a measure of similarity between two lists while preserving the ranking. We also add English (ENG) as an unrelated control language not belonging to the Philippine family tree for comparison. We use the CommonCore readability dataset (Flor et al., 2013) for English as it also has three readability levels, and the level distribution is the most similar to the dataset of the three Philippine languages. Further information on the datasets in Tagalog, Bikol, and Cebuano can be found in Section 3. For all languages, we extract the top  $25\%$  of the most frequently occurring bigrams and trigrams for analysis. The top 40 most frequent bigrams and trigrams can be found in the Appendix.
+
+Table 1 presents character overlap for bigrams and trigrams in a pairwise manner. These results show that all three Philippine languages have character overlap greater than  $75\%$  for bigrams among themselves while overlap with English is below  $27\%$ . This pattern is observed again
+
+in trigrams with the overlap levels of  $53.3\%$  to  $62.8\%$  between Tagalog, Bikol, and Cebuano and those below  $15\%$  for English. These ranges of mutual intelligibility values for bigram and trigram overlap serve as an estimate of the degree of relatedness between the three Philippine languages, with the values for English serving as a baseline for an unrelated language.
+
+Genetic Distance. As a secondary measure of mutual intelligibility, we calculate the genetic distance score (Beau and Crabbé, 2022) for each pair of languages studied in this work. Similar to the character n-gram overlap analysis, we add English for comparison purposes. Genetic distance (Beau-fils and Tomin, 2020) is an automatic measure for quantifying the distance between two languages without the need for human judgments. This metric requires a list of words and their equivalent translations for any two languages of interest and calculates the number of exact consonant matches using the following formula:
+
+$$
+\mathrm {G e n e t i c D i s t a n c e} = 1 0 0 - \left(\frac {\operatorname {m a t c h} \left(l _ {1} , l _ {2}\right)}{n}\right) (1)
+$$
+
+where  $l_{1}$  and  $l_{2}$  are a pair of languages,  $n$  is the total number of words for analysis (usually 100), and match(·) is a function for extracting the consonant patterns for each word from the list as described in Beaufils and Tomin (2020). The metric is measured as a distance; thus, the values closer to 100 denote higher dissimilarity or non-relatedness.
+
+(b) Trigram Character Overlap  
+Table 1: Mutual Intelligibility using Bigram and TriGram Character N-Gram Overlap  
+
+<table><tr><td></td><td>TGL</td><td>BCL</td><td>CEB</td><td>ENG</td></tr><tr><td>TGL</td><td>0.000</td><td>37.083</td><td>24.846</td><td>95.690</td></tr><tr><td>BCL</td><td>37.083</td><td>0.000</td><td>31.933</td><td>70.735</td></tr><tr><td>CEB</td><td>24.846</td><td>31.933</td><td>0.000</td><td>90.970</td></tr><tr><td>ENG</td><td>95.690</td><td>70.735</td><td>90.970</td><td>0.000</td></tr></table>
+
+<table><tr><td>Range</td><td>Meaning</td></tr><tr><td>Between 1 and 30</td><td>Highly related languages</td></tr><tr><td>Between 30 and 50</td><td>Related languages</td></tr><tr><td>Between 50 and 70</td><td>Remotely related languages</td></tr><tr><td>Between 70 and 78</td><td>Very remotely related languages</td></tr><tr><td>Between 78 and 100</td><td>No recognizable relationship</td></tr></table>
+
+Table 2: Mutual Intelligibility using Genetic Distance with Mapping from Beaufils and Tomin (2020).
+
+Table 2 shows the calculated genetic distance scores for each pair of languages including English. The mapping provided in the table is the prescribed guide from Beaufils and Tomin (2020).
+
+Judging by these results, the Philippine languages have genetic distance scores within the related and highly related languages range with the Tagalog-Cebuano pair showing the closest language distance of 24.846. Meanwhile, genetic distance scores between all considered Philippine languages and English fall within the very remotely related to no recognizable relationship categories, with the Tagalog-English pair showing the highest distance from each other. Similar to the character n-gram overlap, these results strengthen our initial observation and provide empirical evidence for mutual intelligibility between Tagalog, Bikol, and Cebuano languages which, beyond this study, may also be used in future linguistic research.
+
+# 3 Readability Corpora in Philippine Languages
+
+We have compiled open source readability datasets for Tagalog, Cebuano, and Bikol from online library websites and repositories. Each data instance in this study is a fictional short story. Table 3 shows the statistical breakdown and additional information on the levels in each readability dataset across different languages.
+
+Tagalog and Cebuano. Datasets in these languages have already been used in previous research, including Imperial et al. (2019); Imperial and Ong (2020); Imperial (2021); Imperial and Ong (2021a); Imperial et al. (2022). We use the same datasets as in previous research and incorporate them into this study for comparison. For Tagalog, we have assembled 265 instances of children's fictional stories from Adarna House<sup>3</sup> and the Department of Education (DepED)<sup>4</sup>. For Cebuano, we use the dataset collected by Imperial et al. (2022) from Let's Read Asia<sup>5</sup> and Bloom Library<sup>6</sup>, which were funded by the Summer Institute of Linguistics (SIL International) and BookLabs to make literary materials in multiple languages available to the public.
+
+Bikol. There are no pre-compiled datasets available for readability assessment in Bikol yet. For this, we collected all available Bikol short stories from Let's Read Asia and Bloom Library totaling 150 instances split into 68, 27, and 55 for
+
+levels 1 to 3 respectively.
+
+All collected data for this study follows the standard leveling scheme for early-grade learners or the first three grades from the K-12 Basic Curriculum in the Philippines. Each instance has been annotated by experts with a level from 1 to 3 as seen in Table 3. We use these annotations as target labels in our experiments. Finally, all datasets used in this study can be manually downloaded from their respective websites (see footnotes for links) under the Creative Commons BY 4.0 license.
+
+# 4 Experimental Setup
+
+# 4.1 ML Setup
+
+In this study, our primary focus is on the depth of analysis of the traditional and neural features used in a cross-lingual setting applied to closely related languages. Thus, we use a vanilla Random Forest model which has been previously shown to be the best-performing monolingual-trained model for ARA in Tagalog and Cebuano (Imperial and Ong, 2021a; Imperial et al., 2022). We leave the technical breadth of exploring other supervised algorithms to future work.
+
+We use a stratified  $k$ -fold approach with  $k = 5$  to have well-represented samples per class for a small-dataset scenario used in this study. We report accuracy as the main evaluation metric across all experiments for the ease of performance comparison with previous work (see Section 5). We use WEKA 3.8 (Witten et al., 1999)<sup>8</sup> for all our modeling and evaluation and set hyperparameters of the Random Forest algorithm to their default values as listed in the Appendix.
+
+# 4.2 Linguistic Features
+
+We extract and consider a wide variety of features inspired by: (a) handcrafted predictors from previous work, (b) representations from a multilingual Transformer-based model (mBERT), and (c) CROSSNGO, a novel feature applicable to readability assessment in closely related languages. We discuss each feature group below.
+
+Traditional Handcrafted Features (TRAD). We integrate available traditional surface-based and
+
+<table><tr><td>Source</td><td>Language</td><td>Level</td><td>Doc Count</td><td>Sent Count</td><td>Vocab</td></tr><tr><td rowspan="3">Adarna and DepED</td><td rowspan="3">TGL (265)</td><td>L1</td><td>72</td><td>2774</td><td>4027</td></tr><tr><td>L2</td><td>96</td><td>4520</td><td>7285</td></tr><tr><td>L3</td><td>97</td><td>10957</td><td>12130</td></tr><tr><td rowspan="3">Let&#x27;s Read Asia and Bloom Library</td><td rowspan="3">BCL (150)</td><td>L1</td><td>68</td><td>1578</td><td>2674</td></tr><tr><td>L2</td><td>27</td><td>1144</td><td>2009</td></tr><tr><td>L3</td><td>55</td><td>3347</td><td>5509</td></tr><tr><td rowspan="3">Let&#x27;s Read Asia and Bloom Library</td><td rowspan="3">CBL (349)</td><td>L1</td><td>167</td><td>1173</td><td>2184</td></tr><tr><td>L2</td><td>100</td><td>2803</td><td>4003</td></tr><tr><td>L3</td><td>82</td><td>3794</td><td>6115</td></tr></table>
+
+Table 3: Statistics on the readability corpora in Tagalog, Cebuano, and Bikol used in this study. The numbers in the brackets provided in the second column are the total number of documents per language broken down in the third and fourth columns per grade level. Doc Count and Sent Count denote the number of short story instances and the number of sentences per story. Vocabulary is the size of the vocabulary or of the accumulated unique word lists per level.
+
+syllable pattern-based features in this study as predictors of text complexity. These features have been widely used in previous research on ARA in Tagalog and Cebuano (Imperial and Ong, 2020; Imperial et al., 2022). For Bikol, this is the first-ever study to develop a readability assessment model. In the case of low resource languages similar to those used in this study, these predictors are still the go-to features in ARA and have been empirically proven effective for Tagalog and Cebuano (Imperial and Ong, 2021b). We have extracted a total of 18 traditional features for each language, including:
+
+1. The total number of words, phrases, and sentences (3).  
+2. Average word length, sentence length, and the number of syllables per word (3).  
+3. The total number of polysyllable words of more than 5 syllables (1).  
+4. Density of consonant clusters or frequency of consonants without intervening vowels in a word (e.g. Tagalog: sastre, En: dressmaker) (1).  
+5. Densities of syllable patterns using the following templates  $\{\mathrm{v},\mathrm{cv},\mathrm{vc},\mathrm{cvc},\mathrm{vcc},\mathrm{ccv},\mathrm{cvcc},$  ccvc, ccvcc, ccvccc}, where v and c are vowels and consonants respectively (10).
+
+Multilingual Neural Embeddings (mBERT). In addition to the surface-based features, we explore contextual representations from a multilingual Transformer-based large language model via mBERT (Devlin et al., 2019). Previous research on probing BERT has shown convincing evidence
+
+that various types of linguistic information (e.g. semantic and syntactic knowledge) are distributed within its twelve layers (Tenney et al., 2019; Rogers et al., 2020). Applying this to ARA, Imperial (2021) showed that BERT embeddings could act as a substitute feature set for lower-resource languages such as Filipino, for which NLP tools like POS taggers are lacking.
+
+For this study, we specifically chose mBERT as this particular model has been trained using Wikipedia data in 104 different languages including Tagalog and Cebuano. Bikol is not included in any available off-the-shelf Transformer-based language models due to extremely limited online resources not large enough for training. Nonetheless, we still used the representations provided by mBERT noting its high intelligibility with Tagalog and Cebuano. Feature-wise, we use the mean-pooled representations of the entire twelve layers of mBERT via the sentence-transformers library (Reimers and Gurevych, 2019). Each instance in our readability data has an mBERT embedding representation of 768 dimensions.
+
+Cross-lingual Character N-Gram Overlap (CROSSNGO). N-gram overlap has been used previously in various NLP tasks applied to Philippine language data such as language identification (Oco et al., 2013a; Cruz et al., 2016), spell checking and correction (Cheng et al., 2007; Octaviano and Borra, 2017; Go et al., 2017), and clustering (Oco et al., 2013b). Drawing inspiration from this fact and from the quantitative evidence of mutual intelligibility between Philippine languages presented in Section 2, we posit that a new feature designed specifically for closely related language data might improve the performance of the readability assess
+
+ment models. Thus, we introduce CROSSNGO, which quantifies linguistic similarity using character overlap from a curated list of high-frequency n-grams within languages of high mutual intelligibility. We propose the following formula for calculating this metric:
+
+$$
+\operatorname {C r o s s N G O} _ {L, n} = \frac {m (L) \cap m (d)}{\operatorname {c o u n t} (m (d))} \tag {2}
+$$
+
+where  $n \in \{2,3\}$  denotes bigrams and trigrams, and  $m(\cdot)$  is a function that extracts unique n-grams from a document instance  $d$  and compares them to a list of top n-grams from a specific language  $L$ . For each instance in a dataset, a vector containing three new features will be added representing the overlap between the text and the top n-grams from each of the three languages. We apply this calculation to both bigrams and trigrams using the n-gram lists for Tagalog, Bikol, and Cebuano obtained from the preliminary experiments, which results in a total of 6 new features.
+
+While we presented two quantitative methods of mutual intelligibility in Section 2, only CROSS-NGO is applied as a metric and a feature for this study. Staying faithful to the work of Beaufils and Tomin (2020), we did not use Genetic Distance to generate another set of features as it was originally developed as a language-to-language metric. Thus, we use it only as additional secondary evidence of language similarity. At the same time, we note that the proposed CROSS-NGO bears certain conceptual similarities to Genetic Distance as it measures the frequency of n-gram overlap with other languages. We perform an ablation study and demonstrate the contribution of individual feature sets in Section 5.
+
+# 5 Results and Discussion
+
+Table 4 shows the accuracy values obtained when training Random Forest models with various combinations of feature groups for each language of interest. The experiments were divided into three setups: (a) singular cross-lingual  $(l_{1}\rightarrow l_{2})$ , (b) pairwise cross-lingual  $([l_{1} + l_{2}]\rightarrow l_{3})$ , and (c) full cross-lingual  $([l_{1} + l_{2} + l_{3}]\rightarrow l_{1})$ , each corresponding to a separate subsection of Table 4. We use the term cross-lingual in this context when a model is trained with a readability corpus from a chosen language  $l_{n}$  or a combination of languages and evaluated with a test set from another language  $l_{m}$  as is demonstrated in Table 4. Similar to our preliminary experiments (Section 2), we include English using
+
+the CommonCore dataset as counter-evidence for comparison with closely related languages.
+
+# 5.1 Low-Resource Languages Benefit from Specialized Cross-lingual Features
+
+For the singular cross-lingual experiments, the effectiveness of exploiting the bigram and trigram overlap via CROSSNGO is demonstrated by high scores for Bikol and Cebuano (75.862 and 78.270) and comparable performance for Tagalog (50.100). Moreover, only for this setup, there is an observed trend where traditional features combined with CROSSNGO outperform mBERT embeddings or the combination of all features for the respective language pair  $l_{1} \rightarrow l_{2}$ . For Tagalog, this results in 50.100 vs. 26.921 and 23.077; for Bikol - 75.862 vs. 68.965 and 69.000; for Cebuano - 78.270 vs. 71.015 and 73.913. In terms of cross-linguality, in the case of Tagalog, using a model trained with Bikol data proves to be more effective than training with the original Tagalog data with approximately 5.8-point difference in accuracy. However, we still recommend the Tagalog model using all features with 50.000 accuracy since the 0.1 difference is not a significant improvement. Consequently, this trend is not observed in the Bikol and Cebuano experiments where the best-performing models of readability assessment are trained on the data from the same language  $l_{1} \rightarrow l_{1}$ .
+
+To further confirm if the addition of the CROSS-NGO feature statistically improves models' performance as compared to the representations from mBERT for low-resource languages, we aggregate the scores from the TRAD+CROSSNGO group and compare them with the scores obtained when we use mBERT embeddings only, conducting a  $t$ -test. We did not include the scores using the combination of all types of features as it would confound the significance test. We achieve statistical significance at  $\alpha = 0.01$  level ( $p = 0.006$ ) which shows that using traditional handcrafted features extended with CROSSNGO significantly improves ARA models for low-resource languages, provided the availability of data in a closely related language in the case of non-availability of multilingual LLMs (e.g., lack of mBERT model in Bikol).
+
+# 5.2 Inclusion of a Closely Related Language in Data Produces More Confident Predictions
+
+For pairwise cross-lingual experiments, we investigate the effect of adding a closely related language
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">TGL</td><td colspan="4">BCL</td><td colspan="4">CEB</td></tr><tr><td>TRAD</td><td>TRAD + CrossNGO</td><td>mBERT Emdlng</td><td>ALL</td><td>TRAD</td><td>TRAD + CrossNGO</td><td>mBERT Emdlng</td><td>ALL</td><td>TRAD</td><td>TRAD + CrossNGO</td><td>mBERT Emdlng</td><td>ALL</td></tr><tr><td>TGL</td><td>43.153</td><td>44.231</td><td>46.100</td><td>50.000</td><td>55.172</td><td>41.379</td><td>20.689</td><td>24.137</td><td>53.623</td><td>57.971</td><td>47.826</td><td>50.725</td></tr><tr><td>BCL</td><td>50.000</td><td>50.100</td><td>26.921</td><td>23.077</td><td>74.620</td><td>75.862</td><td>68.965</td><td>69.000</td><td>63.768</td><td>62.320</td><td>60.869</td><td>66.667</td></tr><tr><td>CEB</td><td>32.692</td><td>38.462</td><td>34.615</td><td>42.308</td><td>51.720</td><td>65.517</td><td>48.276</td><td>44.823</td><td>74.058</td><td>78.270</td><td>71.015</td><td>73.913</td></tr><tr><td>ENG*</td><td>26.923</td><td>44.230</td><td>28.846</td><td>26.923</td><td>48.275</td><td>37.681</td><td>48.250</td><td>48.275</td><td>46.375</td><td>62.018</td><td>43.478</td><td>43.376</td></tr><tr><td>TGL+BCL</td><td>51.101</td><td>51.923</td><td>40.384</td><td>57.692</td><td>72.441</td><td>69.965</td><td>69.000</td><td>68.966</td><td>56.521</td><td>60.869</td><td>62.318</td><td>69.565</td></tr><tr><td>BCL+CEB</td><td>48.077</td><td>50.000</td><td>42.307</td><td>48.076</td><td>68.956</td><td>72.414</td><td>75.611</td><td>75.862</td><td>74.400</td><td>75.362</td><td>75.362</td><td>79.710</td></tr><tr><td>CEB+TGL</td><td>44.230</td><td>36.538</td><td>48.076</td><td>48.100</td><td>52.720</td><td>55.172</td><td>41.379</td><td>34.483</td><td>77.711</td><td>76.811</td><td>73.913</td><td>74.464</td></tr><tr><td>ALL</td><td>50.000</td><td>52.910</td><td>46.153</td><td>32.692</td><td>72.413</td><td>79.113</td><td>65.517</td><td>79.328</td><td>77.710</td><td>78.000</td><td>78.261</td><td>75.630</td></tr></table>
+
+Table 4: The accuracy of cross-lingual modeling per language with various iterations using different combinations of traditional and neural-based features. The underlined values correspond to the best model for each of the three setups while the boldfaced values correspond to the overall highest-performing model for each language across all setups. We included English as counter-evidence only for the singular cross-lingual setup.
+
+on a model's performance using confusion matrices. As the middle section of Table 4 demonstrates, there are three possible pairwise combinations of Tagalog, Bikol, and Cebuano tested on each individual language. As there can be numerous ways to analyze the table, we highlight the results of the cross-lingual models with the top-performing pair and their utilized feature groups and compare them to their equivalent models in the singular cross-lingual experiment. Figure 2 illustrates this method of comparison for each language.
+
+In the case of the Tagalog-Tagalog pair, most misclassifications occur between grades 1 and 2 in both training and test data using all features. This, in turn, is alleviated by incorporating the Bikol dataset in the training data, which reduces the level of confusion by approximately  $7\%$ . The inclusion of Bikol also improves classification between grades 2 and 3 by three instances. In the case of the Bikol test data, the same finding is observed for the combined Bikol and Cebuano model using all features, where confusion in classifying grades 1 and 3 is reduced by two instances. Lastly, for Cebuano, the top-performing model in the pairwise cross-lingual setup includes Bikol data and uses all features. For this model, misclassifications in predicting grade 1 against the other two levels are reduced, and performance for predicting grade 3 is improved.
+
+We further corroborate our observations that pairwise cross-lingual models outperform singular cross-lingual models by aggregating the scores from the two setups and running a  $t$ -test. Further to the results reported in the previous section, we observe statistically significant difference at the  $\alpha = 0.01$  level ( $p = 0.003$ ) when pairwise cross-lingual models are compared to singular cross
+
+lingual models. Overall, our findings provide solid empirical evidence that including a closely related language in the training data for a low-resource language significantly improves performance.
+
+# 5.3 Combining Specialized Cross-Linguial Features with Multilingual Neural Embeddings Achieves SOTA Results
+
+While the previous sections highlight the significant increase in performance when using traditional features with CROSSNGO as compared to mBERT embeddings only, we now discuss results and contributions when both linguistic representations are combined. As is demonstrated in Table 4, the scores obtained using the combined features applied to Tagalog and Cebuano achieve state-of-the-art results for ARA in these languages. For Tagalog, our model's accuracy of 57.692 outperforms the SVM with 57.10 accuracy and the Random Forest model with 46.70 presented in Imperial (2021). For Cebuano, our model achieves 79.710 beating the Random Forest model presented in Imperial et al. (2022) with a score of 57.485 with both models utilizing the same Cebuano dataset. Lastly, as there are no automated readability assessment models yet for Bikol, we report a baseline accuracy of 79.328, which is achieved using a model with a combination of traditional features (extended with CROSSNGO) and mBERT embeddings extracted from data in all three Philippine languages.
+
+# 5.4 Conventional Fine-Tuning of mBERT Underperforms for Low Resource Cross-Lingual ARA
+
+While the main focus of our work is on using traditional machine learning models with Random Forest, we explore if the standard approach for fine
+
+![](images/f11d926050fca5af46c5a1ed854ab1b6a36cb69a0b0e953245a11cc7fa031280.jpg)  
+(a) TGL only trained model (left) against TGL+BCL model (right) using all features for ARA in TGL.
+
+![](images/717517916ba9fdeb607989f9c8c3f317959bf83387bc79c204b351447fd46671.jpg)
+
+![](images/04af004c677627840dde502f76e3631486bae6e86ac4ba92e30745d48d930a53.jpg)  
+(b) BCL only trained model (left) against BCL+CEB model (right) using all features for ARA in BCL.
+
+![](images/1086f871c35a42713659f025143652f1f1711efad560a46d64700a9c4ec1c8cd.jpg)
+
+![](images/4ea4d20192076b47a34e4ace9432cf7165a123145d278d44dd0318b49af16c13.jpg)  
+(c) BCL only trained model (left) against BCL+CEB model (right) using multilingual embeddings for ARA in CEB.
+
+![](images/e06d2504061c6da8565e4c7f78eb7d9480d939c24b3a20a7d0b29e6ce27b7410.jpg)  
+Figure 2: Confusion matrices for pairwise cross-lingual setup for all three languages. All models using an additional language dataset for ARA achieved an improved performance with an average of 10.131 across the board (7.692 for TGL, 6.862 for BCL, and 15.841 for CEB).
+
+<table><tr><td>Model</td><td>TGL</td><td>BCL</td><td>CEB</td></tr><tr><td>TGL</td><td>0.420</td><td>0.500</td><td>0.333</td></tr><tr><td>BCL</td><td>0.420</td><td>0.633</td><td>0.575</td></tr><tr><td>CEB</td><td>0.520</td><td>0.500</td><td>0.697</td></tr><tr><td>TGL+BCL</td><td>0.440</td><td>0.566</td><td>0.469</td></tr><tr><td>BCL+CEB</td><td>0.400</td><td>0.637</td><td>0.666</td></tr><tr><td>CEB+TGL</td><td>0.480</td><td>0.500</td><td>0.590</td></tr><tr><td>*ALL</td><td>0.460</td><td>0.633</td><td>0.636</td></tr></table>
+
+Table 5: The accuracy scores of the conventional finetuning strategy applied to LLMs for various methods of cross-lingual ARA using the same uncased mBERT model for the extraction of embeddings.
+
+tuning LLMs such as mBERT can produce comparable performance. We use the same uncased mBERT model as presented in Section 4.
+
+Table 5 shows the performance of singular, pairwise, and full cross-lingual setups formatted similarly to Table 4. These results confirm the findings of Ibañez et al. (2022), who have applied a similar setup to monolingual Tagalog ARA using a Tagalog BERT model. Judging by their results, the conventional fine-tuning approach proved to be inferior to the traditional way of extracting linguistic features from text and training a machine learning model like SVM or Random Forest. For this study, the highest-performing setups for Tagalog and Cebuano use Cebuano data only, and that for Bikol uses the combined Cebuano + Bikol datasets. None of the fine-tuned models outperform those presented in Table 4 using combinations of traditional features and CROSSNGO. While previous work in cross-lingual ARA by Lee and Vajjala (2022) and Madrazo Azpiazu and Pera (2020) achieved relatively high performance with non-closely related languages using LLMs, we obtain less promising results which we can attribute to: (a) the use of datasets of substantially smaller sizes (a total of 13,786 documents used in Azpiazu and Pera (2019) and 17,518 in Lee and Vajjala (2022) vs. only 764 in our study), and (b) lack of diverse data sources since only Wikipedia dumps were used for Tagalog and Cebuano for training the mBERT model.
+
+# 6 Conclusion
+
+In this work, we took a step back from the trend of exploring various technical components of the complex, deep learning models and, instead, focused on studying the potential effectiveness of linguistic characteristics such as mutual intelligibility for ARA in closely related Philippine languages – Tagalog, Bikol, and Cebuano. We implemented three cross-lingual setups to closely study the effects of interaction between the three languages and proposed a new feature utilizing n-gram overlap, CROSSNGO, which is specially developed for cross-lingual ARA using closely related languages. Our results show that: (a) using CROSSNGO combined with handcrafted features achieves significantly higher performance than using mBERT embeddings, (b) the inclusion of another closely related Philippine language reduces model confusion, and (c) using the conventional fine-tuning for LLMs like mBERT in this setup still does not
+
+outperform models with traditional features. Consequently, we come to the conclusion that using languages with high intelligibility is more suited for cross-lingual ARA. This is demonstrated in experiments with English added as an example of a non-related language, in which we do not achieve a substantial increase in performances for Tagalog, Cebuano, and Bikol.
+
+Our results agree with the findings of previous studies in cross-lingual ARA such as those of Madrazo Azpiazu and Pera (2020) using English, Spanish, Basque, Italian, French, Catalan, and Weiss et al. (2021) using English and German, that also showed that the inclusion of additional language data can improve ARA results on other languages. However, our work is primarily motivated by the degree of language relatedness: we show that better results can be achieved for ARA in low-resource languages if we use closely related languages rather than any language, including non-related ones like English. Our study also provides an encouragement for researchers to consider approaches grounded in linguistic theories which can potentially be used to improve the performance in NLP tasks rather than always resorting to models that are expensive to train and hard to interpret.
+
+# 7 Limitations
+
+We discuss some limitations of our current work which can be further explored in the future.
+
+On Data Format. We specifically use fictional short stories as our primary data for the study since we require gold standard labels for this document classification task. Moreover, fictional short stories are easier to find as they often come with a specified grade level compared to other types of literary texts such as magazines or web articles written in any of the three Philippine languages. We do not claim that our models are able to generalize on these other types of literary materials or on other types of closely related language pairs unless a full study is conducted which is outside the scope of this work.
+
+On Handcrafted Features. We were only able to use traditional handcrafted features covering count-based predictors such as sentence or word count and syllable pattern-based features for training the Random Forest models. We did not extract other feature sets one may find in the previous work on
+
+English such as lexical density or discourse-based features since such features require NLP tools that are able to extract POS, named entities, relations, and discourse patterns that do not yet exist for all three Philippine languages used in this study. The work of Imperial and Ong (2021b) covered a small set of lexical features such as type-token ratio and compound word density for readability assessment in Tagalog. Still, we cannot use this approach since all languages would need to have the same number of features as is a standard practice in model training.
+
+On Model Training. Our choice of the Random Forest algorithm for training the ARA models is based on the substantial amount of previous work supporting the application of this method to low-resource ARA, e.g., to Tagalog and Cebuano in a monolingual setup (Imperial and Ong, 2020, 2021a; Imperial, 2021; Imperial et al., 2022), where it achieved better results than other algorithms such as SVM or Logistic Regression. One can consider these algorithms for comparison but the analysis of each ARA model trained with various algorithms to the same level of depth and focus that we have given to the Random Forest classifier in the present study would require a considerable amount of time as well as a higher page limit.
+
+On Current Measures of Mutual Intelligibility. The majority of existing literature in linguistics, specifically on the topic of mutual intelligibility in Philippine languages, discusses examples in the context of speech communication. As such, one might claim that Cebuano and Tagalog are not mutually intelligible by giving an example where a Tagalog speaker may not fully comprehend (or only recognize a few common words) another speaker if they are talking in Cebuano. While this is certainly true, in this study, we specifically focus on the mutual intelligibility of languages at a word and character level via written texts such as children's fiction books. From this, we see a substantial degree of closeness between Tagalog, Cebuano, and Bikol compared to English. Thus, based on our results, we posit that mutual intelligibility may be used as an additional feature (see CROSSNGO in Section 4) for text-based tasks such as readability assessment. We leave the exploration of our proposed novel feature in the speech communication
+
+area to future work.
+
+# 8 Ethical Considerations
+
+We foresee no ethical issues related to the study.
+
+# Acknowledgements
+
+We thank the anonymous reviewers and area chairs for their constructive and helpful feedback. We also thank the communities and organizations behind the creation of open-source datasets in Philippine languages used in this research: DepED, Adarna House, Bloom Library, Let's Read Asia, SIL, and BookLabs. JMI is supported by the UKRI CDT in Accountable, Responsible, and Transparent AI of the University of Bath and by the Study Grant Program of the National University Philippines.
+
+# References
+
+Ion Madrazo Azpiazu and Maria Soledad Pera. 2019. Multiattentive recurrent neural network architecture for multilingual readability assessment. Transactions of the Association for Computational Linguistics, 7:421-436.  
+Nathanaël Beau and Benoit Crabbé. 2022. The impact of lexical and grammatical processing on generating code from natural language. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 2204–2214, Dublin, Ireland. Association for Computational Linguistics.  
+Vincent Beaufils and Johannes Tomin. 2020. Stochastic approach to worldwide language classification: the signals and the noise towards long-range exploration. SocArXiv.  
+Leonard Bloomfield. 1926. A set of postulates for the science of language. Language, 2(3):153-164.  
+Charibeth Cheng, Cedric Paul Alberto, Ian Anthony Chan, and Vazir Joshua Querol. 2007. SpellChef: spelling checker and corrector for Filipino. Journal of Research in Science, Computing and Engineering, 4(3):75-82.  
+Ernesto A Constantino. 1998. Current topics in Philippine linguistics. In Revised version of the paper read at the meeting of the Linguistic Society of Japan held in Yamaguchi University, Yamaguchi, Japan, on 31 October.  
+Angelica Dela Cruz, Nathaniel Oco, Leif Romeritch Syliongka, and Rachel Edita Roxas. 2016. Phoneme inventory, trigrams and geographic location as features for clustering different philippine languages. In 2016 Conference of The Oriental Chapter of International Committee for Coordination and Standardization of Speech Databases and Assessment Techniques (O-COCOSDA), pages 137-140. IEEE.
+
+Tovly Deutsch, Masoud Jasbi, and Stuart Shieber. 2020. Linguistic features for readability assessment. In Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 1-17, Seattle, WA, USA  $\rightarrow$  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.  
+Michael Flor, Beata Beigman Klebanov, and Kathleen M. Sheehan. 2013. Lexical tightness and text complexity. In Proceedings of the Workshop on Natural Language Processing for Improving Textual Accessibility, pages 29-38, Atlanta, Georgia. Association for Computational Linguistics.  
+Thomas François and Cédrick Fairon. 2012. "AI readability" formula for French as a foreign language. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 466-477.  
+Matthew Phillip Go and Nicco Nocon. 2017. Using Stanford part-of-speech tagger for the morphologically-rich Filipino language. In Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation, pages 81-88. The National University (Phillippines).  
+Matthew Phillip Go, Nicco Nocon, and Allan Borra. 2017. Gramatika: A grammar checker for the low-resourced Filipino language. In TENCON 2017-2017 IEEE Region 10 Conference, pages 471-475. IEEE.  
+Julia Hancke, Sowmya Vajjala, and Detmar Meurers. 2012. Readability classification for German using lexical, syntactic, and morphological features. In Proceedings of COLING 2012, pages 1063-1080, Mumbai, India. The COLING 2012 Organizing Committee.  
+Michael Heilman, Kevyn Collins-Thompson, and Maxine Eskenazi. 2008. An analysis of statistical models and features for reading difficulty prediction. In Proceedings of the Third Workshop on Innovative Use of NLP for Building Educational Applications, pages 71-79, Columbus, Ohio. Association for Computational Linguistics.  
+Michael Ibanez, Lloyd Lois Antonie Reyes, Ranz Sapinit, Mohammed Ahmed Hussien, and Joseph Marvin Imperial. 2022. On Applicability of Neural Language Models for Readability Assessment in Filipino. In International Conference on Artificial Intelligence in Education, pages 573-576. Springer.
+
+Joseph Marvin Imperial. 2021. BERT embeddings for automatic readability assessment. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 611-618, Held Online. INCOMA Ltd.  
+Joseph Marvin Imperial and Ethel Ong. 2020. Exploring hybrid linguistic feature sets to measure Filipino text readability. In 2020 International Conference on Asian Language Processing (IALP), pages 175-180. IEEE.  
+Joseph Marvin Imperial and Ethel Ong. 2021a. *Diverse linguistic features for assessing reading difficulty of educational Filipino texts*. arXiv preprint arXiv:2108.00241.  
+Joseph Marvin Imperial and Ethel Ong. 2021b. Under the microscope: Interpreting readability assessment models for Filipino. In Proceedings of the 35th Pacific Asia Conference on Language, Information and Computation, pages 1-10, Shanghai, China. Association for Computational Linguistics.  
+Joseph Marvin Imperial, Lloyd Lois Antonie Reyes, Michael Antonio Ibanez, Ranz Sapinit, and Mohammed Hussien. 2022. A baseline readability model for Cebuano. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 27-32, Seattle, Washington. Association for Computational Linguistics.  
+Joseph Marvin Imperial, Rachel Edita Roxas, Erica Mae Campos, Jemelee Oandasan, Reyniel Caraballo, Ferry Winsley Sabdani, and Ani Rosa Almaroi. 2019. Developing a machine learning-based grade level classifier for Filipino children's literature. In 2019 International Conference on Asian Language Processing (IALP), pages 413-418. IEEE.  
+Zahurul Islam, Alexander Mehler, and Rashedur Rahman. 2012. Text readability classification of textbooks of a low-resource language. In Proceedings of the 26th Pacific Asia Conference on Language, Information, and Computation, pages 545-553.  
+Zahurul Islam and Rashedur Rahman. 2014. Readability of Bangla news articles for children. In Proceedings of the 28th Pacific Asia Conference on Language, Information and Computing, pages 309-317.  
+Justin Lee and Sowmya Vajjala. 2022. A neural pairwise ranking model for readability assessment. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 3802-3813, Dublin, Ireland. Association for Computational Linguistics.  
+Ion Madrazo Azpiazu and Maria Soledad Pera. 2020. Is cross-lingual readability assessment possible? Journal of the Association for Information Science and Technology, 71(6):644-656.  
+Curtis D McFarland. 2004. The Philippine language situation. World Englishes, 23(1):59-75.
+
+Nathaniel Oco, Joel Ilao, Rachel Edita Roxas, and Leif Romeritch Sylliongka. 2013a. Measuring language similarity using trigrams: Limitations of language identification. In 2013 International Conference on Recent Trends in Information Technology (ICRTIT), pages 478-481. IEEE.  
+Nathaniel Oco, Leif Romeritch Syllongka, Rachel Edita Roxas, and Joel Ilao. 2013b. Dice's coefficient on trigram profiles as metric for language similarity. In 2013 International Conference Oriental COCOSDA held jointly with 2013 Conference on Asian Spoken Language Research and Evaluation (O-COCOSDA/CASLRE), pages 1-4. IEEE.  
+Manolito Octaviano and Allan Borra. 2017. A spell checker for a low-resourced and morphologically rich language. In TENCON 2017-2017 IEEE Region 10 Conference, pages 1853-1856. IEEE.  
+Ildiko Pilán, Sowmya Vajjala, and Elena Volodina. 2016. A readable read: Automatic assessment of language learning materials based on linguistic complexity. arXiv preprint arXiv:1603.08868.  
+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.  
+Robert Reynolds. 2016. Insights from Russian second language readability classification: complexity-dependent training requirements, and feature evaluation of multiple categories. In Proceedings of the 11th Workshop on Innovative Use of NLP for Building Educational Applications, pages 289-300, San Diego, CA. Association for Computational Linguistics.  
+Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics, 8:842-866.  
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics.  
+Sowmya Vajjala and Ivana Lucic. 2018. OneStopEnglish corpus: A new corpus for automatic readability assessment and text simplification. In Proceedings of the Thirteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 297-304, New Orleans, Louisiana. Association for Computational Linguistics.  
+Sowmya Vajjala and Detmar Meurers. 2012. On improving the accuracy of readability classification using insights from second language acquisition. In
+
+Proceedings of the seventh workshop on building educational applications using NLP, pages 163-173.
+
+Sowmya Vajjala and Taraka Rama. 2018. Experiments with universal CEFR classification. In Proceedings of the Thirteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 147-153, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Charles Walton. 1979. A Philippine language tree. Anthropological linguistics, 21(2):70-98.
+
+William Webber, Alistair Moffat, and Justin Zobel. 2010. A similarity measure for indefinite rankings. ACM Transactions on Information Systems (TOIS), 28(4):1-38.
+
+Zarah Weiss, Xiaobin Chen, and Detmar Meurers. 2021. Using broad linguistic complexity modeling for crosslingual readability assessment. In Proceedings of the 10th Workshop on NLP for Computer Assisted Language Learning, pages 38-54, Online. LiU Electronic Press.
+
+Zarah Weiss and Detmar Meurers. 2022. Assessing sentence readability for German language learners with broad linguistic modeling or readability formulas: When do linguistic insights make a difference? In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 141-153, Seattle, Washington. Association for Computational Linguistics.
+
+Ian H Witten, Eibe Frank, Leonard E Trigg, Mark A Hall, Geoffrey Holmes, and Sally Jo Cunningham. 1999. Weka: Practical machine learning tools and techniques with Java implementations.
+
+Menglin Xia, Ekaterina Kochmar, and Ted Briscoe. 2016. Text readability assessment for second language learners. In Proceedings of the 11th Workshop on Innovative Use of NLP for Building Educational Applications, pages 12-22, San Diego, CA. Association for Computational Linguistics.
+
+# A Appendix
+
+<table><tr><td>Hyperparameter</td><td>Value</td></tr><tr><td>batchSize</td><td>100</td></tr><tr><td>bagSizePercent</td><td>100</td></tr><tr><td>maxDepth</td><td>unlimited</td></tr><tr><td>numIterations</td><td>100</td></tr><tr><td>numFeatures</td><td>int(log(#predictors) + 1)</td></tr><tr><td>seed</td><td>1</td></tr></table>
+
+Table 6: Hyperparameter settings for the Random Forest algorithm used for training the models in WEKA. These are default values and the 3.8.6 version of WEKA would have these already preset.  
+
+<table><tr><td>Hyperparameter</td><td>Value</td></tr><tr><td>max seq length</td><td>300</td></tr><tr><td>batch size</td><td>8</td></tr><tr><td>dropout</td><td>0.01</td></tr><tr><td>optimizer</td><td>Adam</td></tr><tr><td>activation</td><td>ReLU</td></tr><tr><td>layer count</td><td>1 (768 x 256)</td></tr><tr><td>loss</td><td>Negative Log Likelihood</td></tr><tr><td>learning rate</td><td>0.002</td></tr><tr><td>epochs</td><td>50</td></tr></table>
+
+Table 7: Hyperparameter settings for the mBERT model used for fine-tuning. Please refer to Ibanez et al. (2022) for more information on these values.
+
+<table><tr><td colspan="2">TGL</td><td colspan="2">CEB</td><td colspan="2">BCL</td></tr><tr><td>bigram</td><td>count</td><td>bigram</td><td>count</td><td>bigram</td><td>count</td></tr><tr><td>ng</td><td>43215</td><td>an</td><td>15636</td><td>an</td><td>12562</td></tr><tr><td>an</td><td>39268</td><td>ng</td><td>14451</td><td>na</td><td>7315</td></tr><tr><td>na</td><td>22041</td><td>sa</td><td>8311</td><td>ng</td><td>6754</td></tr><tr><td>in</td><td>18449</td><td>na</td><td>7167</td><td>in</td><td>6138</td></tr><tr><td>ma</td><td>16501</td><td>ga</td><td>6714</td><td>sa</td><td>5753</td></tr><tr><td>sa</td><td>16037</td><td>ka</td><td>5951</td><td>ka</td><td>5176</td></tr><tr><td>la</td><td>15283</td><td>la</td><td>5638</td><td>ag</td><td>4558</td></tr><tr><td>ka</td><td>14263</td><td>ma</td><td>4889</td><td>ma</td><td>4452</td></tr><tr><td>ag</td><td>12386</td><td>ni</td><td>4701</td><td>on</td><td>3490</td></tr><tr><td>at</td><td>12380</td><td>ta</td><td>4692</td><td>ga</td><td>3462</td></tr><tr><td>pa</td><td>12171</td><td>in</td><td>4591</td><td>pa</td><td>3453</td></tr><tr><td>al</td><td>11521</td><td>pa</td><td>4333</td><td>ni</td><td>3416</td></tr><tr><td>ga</td><td>10818</td><td>ag</td><td>4247</td><td>ak</td><td>3291</td></tr><tr><td>ay</td><td>10771</td><td>on</td><td>4113</td><td>ar</td><td>3012</td></tr><tr><td>ak</td><td>10271</td><td>ay</td><td>3799</td><td>si</td><td>2957</td></tr><tr><td>ni</td><td>9814</td><td>si</td><td>3636</td><td>da</td><td>2920</td></tr><tr><td>ta</td><td>9738</td><td>ya</td><td>3603</td><td>ya</td><td>2886</td></tr><tr><td>si</td><td>9126</td><td>al</td><td>3406</td><td>ta</td><td>2796</td></tr><tr><td>ya</td><td>8724</td><td>at</td><td>3150</td><td>la</td><td>2676</td></tr><tr><td>on</td><td>8288</td><td>ba</td><td>3099</td><td>al</td><td>2658</td></tr><tr><td>ba</td><td>7402</td><td>ak</td><td>3062</td><td>ba</td><td>2613</td></tr><tr><td>it</td><td>7288</td><td>ha</td><td>2729</td><td>ra</td><td>2518</td></tr><tr><td>am</td><td>6667</td><td>iy</td><td>2634</td><td>as</td><td>2447</td></tr><tr><td>iy</td><td>6339</td><td>ug</td><td>2531</td><td>at</td><td>2315</td></tr><tr><td>as</td><td>6210</td><td>il</td><td>2511</td><td>ay</td><td>2187</td></tr><tr><td>ko</td><td>5928</td><td>un</td><td>2502</td><td>ab</td><td>1893</td></tr><tr><td>ha</td><td>5885</td><td>gi</td><td>2460</td><td>ai</td><td>1843</td></tr><tr><td>il</td><td>5857</td><td>li</td><td>2413</td><td>ko</td><td>1840</td></tr><tr><td>ar</td><td>5848</td><td>am</td><td>2327</td><td>ha</td><td>1763</td></tr><tr><td>li</td><td>5696</td><td>ah</td><td>2251</td><td>li</td><td>1697</td></tr><tr><td>ap</td><td>5190</td><td>it</td><td>2059</td><td>ad</td><td>1679</td></tr><tr><td>ab</td><td>5000</td><td>ad</td><td>1834</td><td>ro</td><td>1574</td></tr><tr><td>ra</td><td>4867</td><td>as</td><td>1801</td><td>am</td><td>1544</td></tr><tr><td>da</td><td>4777</td><td>da</td><td>1793</td><td>un</td><td>1316</td></tr><tr><td>aw</td><td>4598</td><td>us</td><td>1781</td><td>ti</td><td>1293</td></tr><tr><td>ti</td><td>4577</td><td>ko</td><td>1771</td><td>nd</td><td>1202</td></tr><tr><td>wa</td><td>4572</td><td>to</td><td>1770</td><td>ap</td><td>1172</td></tr><tr><td>ah</td><td>4410</td><td>aw</td><td>1767</td><td>mg</td><td>1165</td></tr><tr><td>um</td><td>4391</td><td>ab</td><td>1690</td><td>ah</td><td>1164</td></tr><tr><td>bi</td><td>4382</td><td>yo</td><td>1667</td><td>it</td><td>1160</td></tr><tr><td>is</td><td>4286</td><td>ki</td><td>1615</td><td>bi</td><td>1146</td></tr><tr><td>to</td><td>4248</td><td>hi</td><td>1589</td><td>ku</td><td>1140</td></tr><tr><td>mi</td><td>4179</td><td>ap</td><td>1516</td><td>aw</td><td>1139</td></tr><tr><td>un</td><td>4168</td><td>mg</td><td>1504</td><td>wa</td><td>1086</td></tr></table>
+
+Table 8: Full list of the top  $25\%$  bigrams extracted from the Tagalog, Cebuano, and Bikol datasets. The same list is used for calculating overlap via CROSSNGO.
+
+<table><tr><td colspan="2">TGL</td><td colspan="2">CEB</td><td colspan="2">BCL</td></tr><tr><td>trigram</td><td>count</td><td>trigram</td><td>count</td><td>trigram</td><td>count</td></tr><tr><td>ang</td><td>22650</td><td>ang</td><td>7941</td><td>ang</td><td>3350</td></tr><tr><td>ala</td><td>6120</td><td>nga</td><td>3283</td><td>nag</td><td>1721</td></tr><tr><td>ing</td><td>5456</td><td>iya</td><td>2547</td><td>kan</td><td>1518</td></tr><tr><td>ong</td><td>5036</td><td>ing</td><td>1697</td><td>aka</td><td>1507</td></tr><tr><td>iya</td><td>4761</td><td>ala</td><td>1534</td><td>ing</td><td>1434</td></tr><tr><td>lan</td><td>3880</td><td>mga</td><td>1479</td><td>nin</td><td>1389</td></tr><tr><td>ina</td><td>3481</td><td>ila</td><td>1474</td><td>ong</td><td>1374</td></tr><tr><td>aka</td><td>3266</td><td>ana</td><td>1395</td><td>ara</td><td>1210</td></tr><tr><td>nan</td><td>3151</td><td>lan</td><td>1317</td><td>mga</td><td>1164</td></tr><tr><td>ama</td><td>3021</td><td>ong</td><td>1315</td><td>man</td><td>1103</td></tr><tr><td>ara</td><td>3007</td><td>ata</td><td>1306</td><td>yan</td><td>979</td></tr><tr><td>ata</td><td>2976</td><td>usa</td><td>1286</td><td>sin</td><td>947</td></tr><tr><td>ila</td><td>2965</td><td>tan</td><td>1276</td><td>ala</td><td>940</td></tr><tr><td>mga</td><td>2867</td><td>yan</td><td>1172</td><td>iya</td><td>928</td></tr><tr><td>nag</td><td>2797</td><td>han</td><td>1139</td><td>asi</td><td>897</td></tr><tr><td>niy</td><td>2795</td><td>ali</td><td>1061</td><td>sai</td><td>853</td></tr><tr><td>pag</td><td>2793</td><td>nag</td><td>1043</td><td>aba</td><td>835</td></tr><tr><td>yan</td><td>2757</td><td>pag</td><td>982</td><td>ina</td><td>833</td></tr><tr><td>apa</td><td>2716</td><td>aka</td><td>975</td><td>aga</td><td>824</td></tr><tr><td>aga</td><td>2694</td><td>ayo</td><td>933</td><td>ini</td><td>816</td></tr><tr><td>ali</td><td>2622</td><td>aha</td><td>931</td><td>mag</td><td>812</td></tr><tr><td>man</td><td>2574</td><td>nan</td><td>928</td><td>aro</td><td>730</td></tr><tr><td>aha</td><td>2450</td><td>siy</td><td>916</td><td>ako</td><td>730</td></tr><tr><td>uma</td><td>2412</td><td>ako</td><td>868</td><td>gan</td><td>718</td></tr><tr><td>aki</td><td>2376</td><td>pan</td><td>863</td><td>par</td><td>705</td></tr><tr><td>nga</td><td>2281</td><td>ama</td><td>847</td><td>nbs</td><td>702</td></tr><tr><td>mag</td><td>2269</td><td>man</td><td>831</td><td>bsp</td><td>702</td></tr><tr><td>aba</td><td>2253</td><td>ini</td><td>830</td><td>ata</td><td>683</td></tr><tr><td>awa</td><td>2249</td><td>ita</td><td>827</td><td>nga</td><td>683</td></tr><tr><td>kan</td><td>2219</td><td>una</td><td>811</td><td>pag</td><td>639</td></tr><tr><td>tin</td><td>2208</td><td>ina</td><td>763</td><td>ati</td><td>605</td></tr><tr><td>asa</td><td>2142</td><td>aba</td><td>758</td><td>lan</td><td>582</td></tr><tr><td>ako</td><td>2130</td><td>kin</td><td>744</td><td>ion</td><td>576</td></tr><tr><td>hin</td><td>2119</td><td>nak</td><td>727</td><td>nda</td><td>574</td></tr><tr><td>ito</td><td>2033</td><td>ung</td><td>718</td><td>lin</td><td>569</td></tr><tr><td>aya</td><td>2000</td><td>kan</td><td>716</td><td>sak</td><td>567</td></tr><tr><td>ana</td><td>1993</td><td>san</td><td>700</td><td>ano</td><td>553</td></tr><tr><td>gan</td><td>1973</td><td>nah</td><td>700</td><td>ban</td><td>547</td></tr><tr><td>ami</td><td>1934</td><td>ngo</td><td>679</td><td>ind</td><td>538</td></tr><tr><td>san</td><td>1913</td><td>kat</td><td>675</td><td>ron</td><td>530</td></tr><tr><td>nak</td><td>1896</td><td>gan</td><td>665</td><td>apa</td><td>527</td></tr><tr><td>abi</td><td>1878</td><td>ula</td><td>636</td><td>ana</td><td>526</td></tr><tr><td>tan</td><td>1844</td><td>ano</td><td>626</td><td>ili</td><td>524</td></tr><tr><td>siy</td><td>1835</td><td>uot</td><td>611</td><td>ent</td><td>508</td></tr><tr><td>ani</td><td>1773</td><td>ahi</td><td>605</td><td>ada</td><td>502</td></tr></table>
+
+Table 9: Full list of the top  $25\%$  trigrams extracted from the Tagalog, Cebuano, and Bikol datasets. The same list is used for calculating overlap via CROSSNGO.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? Section 7  
+A2. Did you discuss any potential risks of your work? Not applicable. Left blank.  
+A3. Do the abstract and introduction summarize the paper's main claims? Section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+We used WEKA as discussed in Section 4.
+
+B1. Did you cite the creators of artifacts you used? Section 4  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3 for the collected Philippine language data and Section 4 for WEKA  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? The dataset is not scraped from social media sites.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 2-3  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4
+
+C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 5  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 4
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank.

2023/Automatic Readability Assessment for Closely Related Languages/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca6e57977212b923868d908e90cca2c0aae782862807c3f9dfb32ddf47ba7e58
+size 648043

2023/Automatic Table Union Search with Tabular Representation Learning/7244320a-8cb5-4ef7-87f0-c4d2dc4ad4ba_content_list.json

@@ -0,0 +1,2013 @@
+[
+    {
+        "type": "text",
+        "text": "Automatic Table Union Search with Tabular Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            136,
+            90,
+            860,
+            112
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xuming Hu<sup>1*</sup>, Shen Wang<sup>2</sup>, Xiao Qin<sup>2</sup>, Chuan Lei<sup>2</sup>, Zhengyuan Shen<sup>2</sup>, Christos Faloutsos<sup>2</sup>, Asterios Katsifodimos<sup>2</sup>, George Karypis<sup>2</sup>, Lijie Wen<sup>1</sup>, Philip S. Yu<sup>1,3</sup>",
+        "bbox": [
+            117,
+            124,
+            882,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Tsinghua University, $^{2}$ Amazon Web Services, $^{3}$ University of Illinois at Chicago hxml9@mails.tsinghua.edu.cn, wenlj@tsinghua.edu.cn, psyu@uic.edu  $\\{$  shenwa, drxqin, chuanlei, donshen\\} @amazon.com  $\\{$  faloutso, akatsifo, gkarypis\\} @amazon.com",
+        "bbox": [
+            124,
+            159,
+            875,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Given a data lake of tabular data as well as a query table, how can we retrieve all the tables in the data lake that can be unioned with the query table? Table union search constitutes an essential task in data discovery and preparation as it enables data scientists to navigate massive open data repositories. Existing methods identify uniability based on column representations (word surface forms or token embeddings) and column relation represented by column representation similarity. However, the semantic similarity obtained between column representations is often insufficient to reveal latent relational features to describe the column relation between pair of columns and not robust to the table noise. To address these issues, in this paper, we propose a multi-stage self-supervised table union search framework called AUTOTUS, which represents column relation as a vector- column relational representation and learn column relational representation in a multi-stage manner that can better describe column relation for table unionability prediction. In particular, the large language model powered contextualized column relation encoder is updated by adaptive clustering and pseudo label classification iteratively so that the better column relational representation can be learned. Moreover, to improve the robustness of the model against table noises, we propose table noise generator to add table noise to the training table data. Experiments on real-world datasets and synthetic test set augmented with table noise show that AUTOTUS achieves  $5.2\\%$  performance gain over the SOTA baseline.",
+        "bbox": [
+            141,
+            275,
+            460,
+            774
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            782,
+            258,
+            797
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The growing availability of tabular data from academia and industry brings new opportunities for economic growth and social benefit, which has also attracted considerable attention in the natural language processing (NLP) community. A number of",
+        "bbox": [
+            112,
+            806,
+            487,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/f625bcd050011948e72cbcb2112ebd8b84279aee320adb0fa1a9dd6c935bf183.jpg",
+        "table_caption": [
+            "Query Table"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Name</td><td>BP</td><td>Team</td><td>Jersey Number</td><td>CLG</td></tr><tr><td>Kevon Looney</td><td>Milwaukee</td><td>Warriors</td><td>5</td><td>UCUCLA</td></tr><tr><td>Anthony Davis</td><td>Chicago</td><td>Lakers</td><td>3</td><td>KentKentucky</td></tr><tr><td>Devin Booker</td><td>Moss Point</td><td>Suns</td><td>1</td><td>KentKentucky</td></tr></table>",
+        "bbox": [
+            519,
+            260,
+            870,
+            305
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/71c5c01e0c7c8fbbcd1890e5fb932a8703adc76ee5e2ff059b23a52e52f53932.jpg",
+        "table_caption": [
+            "Data Lake Table (Unionable)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Player</td><td>Birth Place</td><td>TM</td><td>POS</td><td>Draft</td></tr><tr><td>Andrew Wiggins</td><td>Toronto</td><td>Golden State Warriors</td><td>Shootung Guard</td><td>R1-1</td></tr><tr><td>Shai Gilgeous-Alexander</td><td>Hamilton</td><td>Oklahoma City Thunder</td><td>Point Gusrd</td><td>R1-11</td></tr><tr><td>Dillon Brooks</td><td>Mississauga</td><td>Memphis Grizzlies</td><td>Small Forward</td><td>R2-45</td></tr></table>",
+        "bbox": [
+            519,
+            315,
+            870,
+            366
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d62e8a5f7b52a5885cf46228c272af899dd68e4ea25ac4bc88648352c821042.jpg",
+        "image_caption": [
+            "Figure 1: The table union search example on Open Data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            367,
+            873,
+            441
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "work has been done, such as Table QA (Chen et al., 2020; Zhu et al., 2021), Table fact verification (Aly et al., 2021; Guo et al., 2022), Table summarization (Xi et al., 2023), Table understanding (Yang et al., 2022; Liu et al., 2022a, 2023), etc. More recently, table union search (Nargesian et al., 2018; Bogatu et al., 2020; Khatiwada and Fan, 2023) is proposed to facilitate tabular dataset discovery applications (Brickley et al., 2019; Galhotra and Khurana, 2020; Santos et al., 2022), which can benefit the tabular data integration and analysis. The table union search aims to find all tables in a data lake that have the columns from the same domain as the query table. With the help of the table union search, the understanding of the tabular data is largely improved and it can potentially benefit other tabular-focused NLP downstream tasks. Therefore, table union search is a non-trivial NLP research problem.",
+        "bbox": [
+            507,
+            480,
+            882,
+            770
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent literature shows that column representation methods are useful for table union search. The basic idea is to map the column to a latent vector space (e.g. column representations (Bogatu et al., 2020; Chepurko et al., 2020) or token embeddings (Zhang and Balog, 2017; Herzig et al., 2020; Fan et al., 2022; Yang et al., 2022)) and then compute the table unionability score based on the column relation obtained from the similarity scores of col",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work done during an internship at Amazon Web Services.",
+        "bbox": [
+            112,
+            892,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3786",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 3786-3800 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            228,
+            945,
+            768,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "unn pairs between query table and target tables. However, these methods only consider the column representation and their similarity as the column relation to compute the table unionability score, which is not enough to describe the relationship between two columns and may cause incorrect prediction. Figure 1 shows an example of the table union search with a query table and a unionable table. By looking at the second columns (in red) of two tables, the computed column similarity score (a scalar) from column embedding is low due to the dissimilar token embeddings. In this case, the column relation is not intuitive, taking only the similarity between column embeddings to describe the column relation is not enough to predict the table unionability accurately. However, if we combine each column representation and model them as an embedding (a vector), the new representation (column relational representation) is more informative and can better describe the unintuitive column relation to make the table unionability prediction more accurately. Therefore, a better approach is needed, which can make the table unionability prediction via the column relational representation.",
+        "bbox": [
+            115,
+            85,
+            489,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "On the other hand, real-world table noises widely exist (Koutras et al., 2021; Liu et al., 2022b). Taking column \"CLG\" in the query table and column \"POS\" in the data lake table as examples, both columns have abbreviated column names, tokens are repeated in column values (e.g., \"UCLA\" becomes \"UCUCLA\") and proximal characters in keyboard are replaced (e.g., \"Shooting Guard\" becomes 'Shootung Guard'). Computing column representations or column relational representations based on token embeddings with tabular noise introduces unavoidable wrong predictions. Therefore, it requires the table union search model to be robust to the table noise during optimization.",
+        "bbox": [
+            115,
+            478,
+            489,
+            702
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address above issues, we propose a multi-stage self-supervised table union search framework called AutotUS, which represents column relation as a vector- column relational representation and learn column relational representation in a multi-stage manner that can better describe column relation for unionability prediction. In particular, the large language model powered contextualized column relation encoder is updated by adaptive clustering and pseudo label classification iteratively so that the better column relational representation can be learned. Moreover, to improve the robustness of the model against table noises, we pro",
+        "bbox": [
+            115,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "pose table noise generator to add table noise to the training table data. Extensive experiments are conducted on three real-world datasets and show significant performance gains compared to existing state-of-the-art methods and demonstrates the robustness of AUTOTUS against the table noise.",
+        "bbox": [
+            512,
+            84,
+            882,
+            179
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are four-folded: 1) We propose a novel framework AUTOTUS that can leverage column relational representation for table union search instead of pairwise column similarity. 2) We propose a multi-stage self-supervised method that can update the large language model powered contextualized column relation encoder in a step by step and iterative way to learn better column relational representation. 3) We propose a table noise generator and add table noise to the training tabular data to improve the robustness of the model against the table noise. 4) We conduct extensive experiment on real-world datasets as well as synthetic test set augmented with table noise and demonstrate significant performance gain of AUTOTUS over the strong baselines and shows the robustness of AUTOTUS against the table noise.",
+        "bbox": [
+            512,
+            181,
+            882,
+            454
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            512,
+            468,
+            707,
+            482
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table union search discovers all tables in a set of data lake tables that are工商联able with the query table. Specifically, we focus on the automatically case where no labeled training data is available. We define a set of data lake tables as  $\\mathcal{T}$ . For each table  $T\\in \\mathcal{T}$  which consists of  $m$  columns  $\\{t_1,t_2,\\dots ,t_m\\}$ . There is a column relation  $e_{ij}\\in \\mathcal{E}$  for each pairs of columns  $t_i,t_j$  that describes their relationship. We employ a contextualized column relational encoder to obtain the column relational representation  $\\mathbf{z}_{e_{ij}}\\in \\mathcal{Z}$  for each column relation  $e_{ij}$ . The table unionability score  $u_{(T_i,T_j)}\\in \\mathcal{U}$  is obtained by the table scorer that goes over all column relational representations between two tables  $T_{i},T_{j}$ . Similar to previous studies (Nargesian et al., 2018; Bogatu et al., 2020; Khatiwada and Fan, 2023), given a query table  $Q$ , table union search retrieves the top- $y$ 工商联able tables  $\\mathcal{Q}\\subseteq \\mathcal{T}$ , where  $|\\mathcal{Q}| = y$  and  $\\forall T\\in \\mathcal{Q}$  and  $T^{\\prime}\\in \\mathcal{T} - \\mathcal{Q}$ , we have  $u_{(Q,T)}\\geq u_{(Q,T^{\\prime})}$ .",
+        "bbox": [
+            512,
+            494,
+            882,
+            818
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            512,
+            828,
+            695,
+            844
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the proposed AU-TOTUS. The framework (as shown in Figure 2) consists of following components: Multi-stage Self-supervised Column Relational Representation",
+        "bbox": [
+            512,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3787",
+        "bbox": [
+            482,
+            928,
+            517,
+            939
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e07cc2860c3430a6eff1d42b3093411838d5ab2f9a7b22466d0cfe22a1f08d2.jpg",
+        "image_caption": [
+            "Figure 2: Overview of AUTOTUS framework."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            136,
+            80,
+            867,
+            252
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Learning, Table Noise Generation and Table Unionability Scorer. We first illustrate the motivation for AUTOTUS: The existing column representation methods use cosine similarity to measure the unionability of column pairs. The column encoder in these methods is usually only trained with a binary classification task, i.e. whether two columns are unionable. In AUTOTUS, the table encoder can benefit from an additional task: the identification of hidden semantic types (domains, task knowledge, etc.) from the column pairs. In the absence of explicit type labels, we use adaptive clustering on column pairs and learn those hidden types. The classification step is then used to update the table encoder, so that the embeddings are generated in a way that minimizes the hidden type prediction errors. In short, the clustering and classification modules learn the hidden unionable semantic types, resulting in a better table encoder for the table union search task.",
+        "bbox": [
+            112,
+            282,
+            489,
+            605
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Multi-stage Self-supervised Column Relational Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            621,
+            443,
+            653
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The Multi-stage Self-supervised Column Relational Representation Learning module aims to extract column relation as column relational representation and learn column relational representation in multi-stage manner that can better describe column relation for unionability prediction. As illustrated in Figure 2, this module consists of three components: Contextualized Column Relation Encoder, Column Relation Encoder Updater via Adaptive Clustering, and Column Relation Encoder Updater via Pseudo Label Classification. The Contextualized Column Relation Encoder serializes and tokenizes the column relations as the input and leverages the pretrained BERT (Devlin et al., 2019) to generate contextualized column relation embeddings. The Column Relation Encoder Updater via",
+        "bbox": [
+            112,
+            661,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Adaptive Clustering optimizes Contextualized Column Relation Encoder through relaxed constraint to obtain better column relational representation and generate the cluster label as the pseudo label. The Column Relation Encoder Updater via Pseudo Label Classification uses the generated pseudo label to further update the Contextualized Column Relation Encoder via more strict constraint to generate improved column relational representation. The above two updates are performed in a bootstrapped loop that iteratively improve the language model to generate better column representations.",
+        "bbox": [
+            507,
+            282,
+            884,
+            476
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Contextualized Column Relation Encoder Contextualized Column Relation Encoder extracts the contextualized column relation embeddings to represent the relational features between two columns. We leverage a large language model, BERT (Devlin et al., 2019), to effectively encode column relations, along with their context tabular information.",
+        "bbox": [
+            507,
+            483,
+            882,
+            611
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The pre-trained large language models (such as BERT) support an input length of up to 512 tokens (Devlin et al., 2019). However, columns in a real-world table may contain thousands of tokens, and randomly discarding tokens will lead to loss of important semantics. Therefore, in this work, we score the importance of each token or cell in a column, and keep cells based on their importance. More specifically, we adopt TF-IDF score to calculate the token importance through its inverse document frequency:  $\\log(|\\mathcal{Y}|)/\\{t|\\text{token} \\in t, t \\in \\mathcal{Y}\\}$ , where  $t$  is a column and  $|\\mathcal{Y}|$  is the number of all data lake columns. Then the cell score is calculated by averaging the TF-IDF of all tokens within the cell. After obtaining the importance score of each column-level cell, the row alignment can ensure the correctness of the table semantics, e.g.: Jimmy Butler's birthplace is aligned to the United States instead of the United Kingdom. Therefore, we cal",
+        "bbox": [
+            507,
+            613,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3788",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "culate the average score of cells in each row and sort them in descending order. Next, we select rows until the total token count reaches the input budget and obtain the table:  $T = [t_{1}, t_{2}, \\dots, t_{m}]$ .",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, we serialize the table by column and follow the marking schema adopted in Soares et al. (2019) to augment  $T$  with two reserved tokens to mark the start and the end of each column. We introduce  $\\langle s \\rangle$  and  $\\langle /s \\rangle$  and inject them into  $T$ :",
+        "bbox": [
+            112,
+            149,
+            489,
+            230
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT = \\left[ \\langle s \\rangle , t _ {1}, \\langle / s \\rangle , \\dots , \\langle s \\rangle , t _ {m}, \\langle / s \\rangle \\right], \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            157,
+            242,
+            489,
+            262
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "as the input token sequence for the encoder.",
+        "bbox": [
+            112,
+            275,
+            440,
+            288
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We denote the Contextualized Column Relation Encoder as  $f_{\\theta}(T,\\langle s\\rangle ,\\langle \\langle s\\rangle \\rangle)$ . Instead of using the output of [CLS] token from BERT itself which summarizes the whole table-level semantics, we get the contextualize column relation semantics by concatenating the representations  $\\langle s\\rangle$  of any two columns in different tables or in the same table, and derive a fixed-length representation  $\\mathbf{z}_{e_{ij}}\\in \\mathbb{R}^{2\\cdot z_R}$ :",
+        "bbox": [
+            112,
+            290,
+            487,
+            420
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {z} _ {e _ {i j}} = \\left[ \\mathbf {c} _ {\\langle s _ {t _ {i}} \\rangle}, \\mathbf {c} _ {\\langle s _ {t _ {j}} \\rangle} \\right]. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            218,
+            432,
+            487,
+            454
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Column Relation Encoder Updater via Adaptive Clustering The Column Relation Encoder Updater via Adaptive Clustering encourages each low-confidence relation label distribution to be closer to the high-confidence cluster centroids, improves the confidence of the clusters, and optimizes Contextualized Column Relation Encoder by relax-constraints to generate better column relational representations. In addition, the cluster label generated for each column relation could be treated as self-supervision signal, which serves as the pseudo label for language model update.",
+        "bbox": [
+            112,
+            466,
+            489,
+            658
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "More specifically, after obtaining  $N$  column relational representations  $\\mathcal{Z} = \\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots,\\mathbf{z}_N\\}$ , Column Relation Encoder Updater via Adaptive Clustering computes  $K$  cluster centroids  $\\{\\xi_k\\in \\mathbb{R}^{h_{AC}}\\}_{k = 1}^K$ , and softly assigns all  $N$  column relation representations to  $K$  clusters. In practice, we first adopt standard  $k$ -means clustering among  $N$  column relation data points  $\\mathbb{R}^{h_{AC}}$  to obtain  $K$  initial cluster centroids  $\\{\\xi_k\\in \\mathbb{R}^{h_{AC}}\\}_{k = 1}^K$ . Inspired by Maaten and Hinton (2008); Hu et al. (2020), we leverage the Student's  $t$ -distribution as the kernel to measure the similarity of the embedding point  $\\mathbf{z}_n$  to each centroid  $\\xi_{k}$  as:",
+        "bbox": [
+            112,
+            659,
+            489,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq _ {n k} = \\frac {\\left(1 + \\left\\| \\mathbf {z} _ {n} - \\xi_ {k} \\right\\| ^ {2} / \\alpha\\right) ^ {- \\frac {\\alpha + 1}{2}}}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\left(1 + \\left\\| \\mathbf {z} _ {n} - \\xi_ {k ^ {\\prime}} \\right\\| ^ {2} / \\alpha\\right) ^ {- \\frac {\\alpha + 1}{2}}}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            132,
+            879,
+            487,
+            923
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $q_{nk}$  can be viewed as the probability of assigning column relation embedding  $\\mathbf{z}_n$  to the cluster  $k$  as the soft assignment and  $\\alpha$  denotes the degree of the Student's  $t$ -distribution. We set  $\\alpha = 1$  for all experiments.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We normalize each cluster by adopting the frequency as an auxiliary target distribution proposed by Xie et al. (2016) in Equation 4 and iteratively refine each cluster with the help of the auxiliary target distribution:",
+        "bbox": [
+            507,
+            165,
+            884,
+            244
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {n k} = \\frac {q _ {n k} ^ {2} / f _ {k}}{\\sum_ {k ^ {\\prime}} q _ {n k ^ {\\prime}} ^ {2} / f _ {k ^ {\\prime}}}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            613,
+            256,
+            882,
+            294
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $f_{k} = \\sum_{n = 1}^{N}q_{nk}, k = 1,2,\\dots ,K$  is the soft cluster frequency. With the auxiliary target distribution, we could encourage each cluster to learn from its high confidence cluster assignments and simultaneously alleviating the bias caused by imbalanced clusters. We define KL divergence loss between the soft assignments  $q_{n}$  and the auxiliary distribution  $p_n$  to train the Contextualized Column Relation Encoder as follows:",
+        "bbox": [
+            507,
+            306,
+            882,
+            451
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {A C} = K L (P | | Q) = \\sum_ {n} \\sum_ {k} p _ {n k} \\log \\frac {p _ {n k}}{q _ {n k}}. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            461,
+            882,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We adopt the label associated with the largest probability as the pseudo label  $\\mu_{n}$  for the  $n$ -th column relation:",
+        "bbox": [
+            507,
+            507,
+            882,
+            552
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mu_ {n} = \\underset {k \\in K} {\\operatorname {a r g m a x}} p _ {n k}. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            621,
+            557,
+            880,
+            581
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In summary, comparing with traditional clustering methods such as  $k$ -means, Column Relation Encoder Updater via Adaptive Clustering adopts an iterative, soft-assignment learning process which encourages high-confidence assignments and uses high-confidence assignments to improve low confidence ones. Therefore, Column Relation Encoder Updater via Adaptive Clustering is robust enough to different cluster numbers  $K$ , thus solving the situation that we have no prior knowledge of specific domains and the relational distributions in the column relational representations. To mitigate the possible negative impact of choosing a unideal initial centroid, Column Relation Encoder Updater via Adaptive Clustering randomly re-selects a set of  $K$  initial centroids if  $\\mathcal{L}_{AC}$  does not decrease after the first epoch.",
+        "bbox": [
+            507,
+            587,
+            882,
+            860
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Column Relation Encoder Updater via Pseudo Label Classification Although the clustering module forces the relational label distribution of",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3789",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "low confidence to be close to the cluster centroids of high confidence to improve the confidence of clustering. However, the clusters may contain some different labels, which affects the clustering purity. Therefore, we utilize the Column Relation Encoder Updater via Pseudo Label Classification module to classify the pseudo labels, and obtain better column relational representations by improving the clustering purity. In practice, the Column Relation Encoder Updater via Adaptive Clustering generates cluster labels  $M = \\{\\mu_1,\\mu_2,\\dots,\\mu_N\\}$  for all column relations as pseudo labels. Column Relation Encoder Updater via Pseudo Label Classification could adopt these pseudo labels as self-supervisions derived from the corpora themselves and use pseudo labels to guide the feature representation learning in Contextualized Column Relation Encoder as well as column relation classifier learning in Column Relation Encoder Updater via Pseudo Label Classification.",
+        "bbox": [
+            110,
+            84,
+            492,
+            405
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The Column Relation Encoder Updater via Pseudo Label Classification module learns to predict the pseudo labels as golden labels. More specifically, we have:",
+        "bbox": [
+            112,
+            406,
+            489,
+            470
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {l} _ {n} = \\tau_ {\\phi} (f _ {\\theta} (T, \\langle s \\rangle , \\langle / s \\rangle)), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            201,
+            481,
+            489,
+            499
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau_{\\phi}$  denotes the column relation classification module with parameters  $\\phi$  and  $\\mathbf{l}_n$  is the probability distribution for the  $n$ -th sample over  $K$  pseudo labels. To find the best performance parameters  $\\theta$  for Contextualized Column Relation Encoder and  $\\phi$  for the classifier, we optimize the classification loss:",
+        "bbox": [
+            112,
+            508,
+            489,
+            619
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {R C} = \\min _ {\\theta , \\phi} \\frac {1}{N} \\sum_ {n = 1} ^ {N} l o s s (\\mathbf {l} _ {n}, \\mathrm {o n e \\_ h o t} (\\mu_ {n})), (8)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            627,
+            487,
+            670
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where loss is the cross entropy loss and one-hot  $(\\mu_{n})$  returns a one-hot pseudo label assignment vector.",
+        "bbox": [
+            112,
+            678,
+            489,
+            727
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2 Table Noise Generator",
+        "text_level": 1,
+        "bbox": [
+            112,
+            737,
+            337,
+            752
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The table noise generator module aims to improve the robustness of the model against the table noise. We firstly summarizes the noise types that exist in the real-world tabular data, and introduce these table noises to the column relations during the self-supervised training. Specifically, we defined two types of table noise: column value noise an column name noise. Here are some representative column value noises: 1) Replace / insert characters with proximal characters, e.g., computer  $\\rightarrow$",
+        "bbox": [
+            112,
+            758,
+            490,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "computer / cocomputer. 2) Delete / repeat characters, e.g., computer  $\\rightarrow$  compuer / computer. 3) Change the numeral display format like scientific notation, e.g.,  $12000\\to 1.2\\mathrm{e}4.$  Here are some representative column name noises: 1) Prefix column names with table name, e.g., traveldestination  $\\rightarrow$  tourism_statistics_traveldestination. 2) Abbreviate column names, e.g., traveldestination  $\\rightarrow$  traDest. 3) Drop vowels, e.g., traveldestination  $\\rightarrow$  trvl_dstntn. 4) Synonym substitution (if exists), e.g., traveldestination  $\\rightarrow$  touristdestination. 5) Acronym, e.g., traveldestination  $\\rightarrow$  t_d.",
+        "bbox": [
+            507,
+            84,
+            884,
+            277
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For each table  $T$  in the data lake tables  $\\mathcal{T}$ , we randomly apply above noise to the training tabular data to improve the robustness of our model against the table noise.",
+        "bbox": [
+            507,
+            279,
+            882,
+            342
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Table Unionability Scorer",
+        "text_level": 1,
+        "bbox": [
+            507,
+            362,
+            757,
+            376
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The Table Scorer obtains the table unionability score based on the column relational representations between column relations in the query table and data lake tables, which can provide more sufficient information to clarify whether the two tables come from the same domain and have a union relation. After we obtained the Contextualized Column Relation Encoder with discriminative power to get the column relation embeddings  $\\mathbf{z}_{e_{ij}} = (\\mathbf{c}_{\\langle s_{t_i}\\rangle},\\mathbf{c}_{\\langle s_{t_j}\\rangle})$  , we adopt the Column Relation Encoder Updater via Pseudo Label Classification to assign the  $K$  relational labels to all  $N$  column relations  $\\mathbf{z}^n = (\\mathbf{z}_{e_{ij}})^n$  in the two tables:",
+        "bbox": [
+            507,
+            386,
+            884,
+            599
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\psi_ {n k} = \\underset {k \\in K, n \\in N} {\\operatorname {a r g m a x}} \\tau_ {\\phi} \\left(\\mathbf {z} ^ {n k}\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            600,
+            615,
+            882,
+            644
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and assign the majority label  $\\kappa$  as the predict relational label for the query table and data lake table. We count the number of majority labels  $\\kappa$  in the all  $N$  column relational representations between two tables as  $M$ , so the table unionability score is  $M / N$ . This score no longer depends on the calculation of cosine similarity between two column features to obtain the column unionability score, but obtains table unionability score through the column unionability scores. The higher the score, the greater the probability of these two tables being unjoinable. Instead of computational methods with insufficient column relational information, AUTO-TUS directly model the complex column relational representations, which can better address the table union search task.",
+        "bbox": [
+            507,
+            661,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3790",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e425d1c23d5daf864c07060cb19fc652f7fe20eda06c75d63ce6fb389eb918a7.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Test / Train</td><td># Tables</td><td># Columns</td><td>Avg. # Rows</td><td>Size (GB)</td></tr><tr><td>SANTOS Small</td><td>550</td><td>6,322</td><td>6,921</td><td>0.45</td></tr><tr><td>TUS Small</td><td>1,530</td><td>14,810</td><td>4,466</td><td>1</td></tr><tr><td>TUS Large</td><td>5,043</td><td>54,923</td><td>1,915</td><td>1.5</td></tr><tr><td>SANTOS Small</td><td>550</td><td>6,162</td><td>7,732</td><td>0.5</td></tr><tr><td>TUS Small</td><td>1,530</td><td>13,926</td><td>4,032</td><td>0.9</td></tr><tr><td>TUS Large</td><td>5,043</td><td>55,462</td><td>1,729</td><td>1.3</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            485,
+            175
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The statistics of datasets for three benchmarks.",
+        "bbox": [
+            112,
+            186,
+            485,
+            200
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Evaluation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            211,
+            366,
+            227
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conduct extensive experiments on real-world public datasets as well as a synthetic test set augmented with table noise to show the effectiveness of our AUTOTUS on table union search task, and give a detailed analysis.",
+        "bbox": [
+            112,
+            236,
+            487,
+            317
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setups",
+        "text_level": 1,
+        "bbox": [
+            112,
+            326,
+            324,
+            343
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. To evaluate our technique, three public datasets are used: SANTOS (Khatiwada and Fan, 2023), TUS small, and TUS large (Nargesian et al., 2018). These three datasets are subsets of Open Data released by authorities such as US Open Data $^{1}$ , UK Open Data $^{2}$ , and Canada Open Data $^{3}$  with manually labeled data. However, these data are usually officially filtered to remove a lot of table noise to maintain high quality. In order to verify the robustness of the model, we synthesized a test set with table noise, we added both column value and column name noises to the tabular data and obtained the synthesized SANTOS $_{N}$  and TUS large $_{N}$ . Note that although we add table noises, the human-annotated labels remain unchanged. To avoid information leakage, the manually labeled dataset is not used during the training. We randomly generated the same number of unlabeled training tables as the labeled tables from Open Data (Koutras et al., 2021; Khatiwada and Fan, 2023). We give the detailed statistics of datasets for three benchmarks in Table 1. We also give the dataset construction details in Appendix A.",
+        "bbox": [
+            112,
+            347,
+            489,
+            718
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Experimental Settings and Metrics. We adopt BERT-Base, BERT-Large (Devlin et al., 2019) and RoBERTa-Large (Liu et al., 2019) as the encoder and set max-length as 512. For Column Relation Encoder Updater via Adaptive Clustering, we adopt  $k$ -means and set  $K = 50$  to get the initial centroids. We stop the clustering and classification loop when the current pseudo labels differs by less than  $10\\%$  from the previous epoch. For Column",
+        "bbox": [
+            112,
+            726,
+            487,
+            871
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Relation Encoder Updater via Pseudo Label Classification, we adopt the fully connected layer as  $\\tau_{\\phi}$  and set dropout rate as  $10\\%$ , learning rate as  $1e - 5$  and warmup as 0.1. To allow fully connected layer to warm up, we fixed the parameters in  $f_{\\theta}$  for the first two epochs. We add one noise of column data or column schemata to each column to obtain the training dataset. We first randomly choose whether to add noise in the column data or column schemata, and secondly, for all replacement, addition, or deletion of noise at the token level, we select  $20\\%$  of all tokens in the cell of the column for modification.",
+        "bbox": [
+            507,
+            84,
+            884,
+            292
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For evaluation metrics, following the previous work (Bogatu et al., 2020; Khatiwada and Fan, 2023), we adopt the Mean Average Precision at  $Y$  ( $MAP@Y$ ) and Recall at  $Y$  ( $R@Y$ ) to evaluate the effectiveness of the searched top- $Y$  table results. Note that  $MAP@Y$  is the average value of Precision at  $y$  ( $P@y$ ), where  $y = 1, 2, \\dots, Y$ . Formally, given the query table  $Q$  and a set of data lake tables  $\\mathcal{T}$ , we define  $\\mathcal{T}_Q$  as the set of usable tables based on the ground truth and  $\\mathcal{T}_Q'$  as the set of top- $Y$ 聢单table results using the searching methods. The  $P@Y$  and  $R@Y$  are calculated as:",
+        "bbox": [
+            507,
+            294,
+            885,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP @ Y = \\frac {\\mathcal {T} _ {Q} \\cap \\mathcal {T} _ {Q} ^ {\\prime}}{\\mathcal {T} _ {Q} ^ {\\prime}}, R @ Y = \\frac {\\mathcal {T} _ {Q} \\cap \\mathcal {T} _ {Q} ^ {\\prime}}{\\mathcal {T} _ {Q}} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            513,
+            882,
+            555
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Note that perfect  $R@Y$  is not possible when the ground truth contains less than  $Y$  searched table results. We define the Mean Average Precision  $MAP@Y$  as:",
+        "bbox": [
+            507,
+            565,
+            882,
+            630
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM A P @ Y = \\frac {1}{Y} \\sum_ {y = 1} ^ {Y} P @ y \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            640,
+            882,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For a fair comparison, we adopt the  $Y = 10$  on the SANTOS benchmarks and  $Y = 60$  on the TUS benchmarks to be consistent with Nargesian et al. (2018); Khatiwada and Fan (2023).",
+        "bbox": [
+            507,
+            699,
+            884,
+            763
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baseline Models. We compare AUTOTUS with baseline models in two categories. The first category is table union search model that adopts various column features to calculate the table unionability score. We adopt  $\\mathbf{D}^3\\mathbf{L}$  (Bogatu et al., 2020), SATO (Zhang et al., 2020), Starmie (Fan et al., 2022), and SANTOS (Khatiwada and Fan, 2023) as baselines. The second category is table pretrained model that leverages unlabeled tables for",
+        "bbox": [
+            507,
+            772,
+            885,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "'https://data.gov/",
+        "bbox": [
+            134,
+            878,
+            302,
+            891
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>2</sup>https://www.data.gov.uk/",
+        "bbox": [
+            134,
+            891,
+            361,
+            904
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://open.canada.ca/en/open-data",
+        "bbox": [
+            136,
+            904,
+            463,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3791",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/9a27ac22b02a66a461cba18922911373d342ec3b3f02b3d1622d385754ee09e3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">SANTOS</td><td colspan=\"2\">TUS small</td><td colspan=\"2\">TUS large</td><td colspan=\"2\">SANTOSN</td><td colspan=\"2\">TUS largeN</td></tr><tr><td>MAP@10</td><td>R@10</td><td>MAP@60</td><td>R@60</td><td>MAP@60</td><td>R@60</td><td>MAP@10</td><td>R@10</td><td>MAP@60</td><td>R@60</td></tr><tr><td>BERT-Base (Devlin et al., 2019)</td><td>79.5</td><td>53.7</td><td>74.1</td><td>16.3</td><td>63.1</td><td>13.4</td><td>69.5</td><td>46.9</td><td>55.3</td><td>9.4</td></tr><tr><td>TaBERT (Yin et al., 2020)</td><td>83.6</td><td>56.2</td><td>80.3</td><td>18.7</td><td>76.4</td><td>15.8</td><td>74.8</td><td>50.7</td><td>67.3</td><td>11.4</td></tr><tr><td>TABBIE (Iida et al., 2021)</td><td>86.3</td><td>58.5</td><td>82.4</td><td>19.2</td><td>78.7</td><td>15.3</td><td>76.9</td><td>51.6</td><td>65.2</td><td>11.8</td></tr><tr><td>TUTA (Wang et al., 2021)</td><td>86.7</td><td>59.2</td><td>84.5</td><td>20.1</td><td>80.5</td><td>16.9</td><td>77.9</td><td>52.3</td><td>71.6</td><td>13.2</td></tr><tr><td>FORTAP (Cheng et al., 2022)</td><td>88.5</td><td>60.9</td><td>87.3</td><td>20.6</td><td>81.3</td><td>17.0</td><td>78.7</td><td>52.7</td><td>72.8</td><td>12.8</td></tr><tr><td>TableFormer (Yang et al., 2022)</td><td>89.4</td><td>61.4</td><td>86.9</td><td>20.2</td><td>83.2</td><td>17.3</td><td>79.2</td><td>53.5</td><td>72.7</td><td>12.9</td></tr><tr><td>D3L (Bogatu et al., 2020)</td><td>82.8</td><td>56.3</td><td>79.5</td><td>19.7</td><td>66.3</td><td>15.2</td><td>73.5</td><td>51.0</td><td>57.8</td><td>9.7</td></tr><tr><td>SATO (Zhang et al., 2020)</td><td>92.7</td><td>68.2</td><td>91.7</td><td>21.5</td><td>86.7</td><td>18.4</td><td>80.1</td><td>54.5</td><td>77.3</td><td>14.5</td></tr><tr><td>Starmie† (Fan et al., 2022)</td><td>93.9</td><td>69.0</td><td>94.8</td><td>23.3</td><td>89.5</td><td>18.9</td><td>81.0</td><td>55.1</td><td>78.4</td><td>14.3</td></tr><tr><td>SANTOS(Khatiwada and Fan, 2023)</td><td>94.3</td><td>69.2</td><td>85.7</td><td>21.0</td><td>-</td><td>-</td><td>81.4</td><td>55.3</td><td>-</td><td>-</td></tr><tr><td>AUTOTUSBERT-Base</td><td>98.5±0.4</td><td>72.4±0.2</td><td>97.7±0.5</td><td>25.2±0.3</td><td>94.6±0.7</td><td>21.6±0.2</td><td>87.6±0.5</td><td>59.7±0.3</td><td>84.9±0.6</td><td>17.3±0.3</td></tr><tr><td>w/o Table Noise Generator</td><td>96.6±0.5</td><td>71.2±0.4</td><td>96.1±0.6</td><td>24.5±0.3</td><td>92.5±0.7</td><td>20.8±0.4</td><td>83.0±0.7</td><td>56.4±0.4</td><td>80.7±0.8</td><td>14.9±0.3</td></tr><tr><td>w/o Encoder Update via Adaptive Clustering</td><td>94.5±0.6</td><td>71.9±0.4</td><td>94.3±0.7</td><td>23.3±0.5</td><td>91.0±0.5</td><td>20.4±0.3</td><td>83.2±0.6</td><td>56.3±0.4</td><td>80.5±0.7</td><td>14.7±0.3</td></tr><tr><td>w/o Encoder Update via Pseudo Label Classification</td><td>81.8±0.6</td><td>54.7±0.4</td><td>80.9±0.9</td><td>16.8±0.4</td><td>77.3±0.7</td><td>17.0±0.4</td><td>72.1±0.8</td><td>50.2±0.6</td><td>68.5±0.7</td><td>12.2±0.2</td></tr><tr><td>AUTOTUSBERT-Large</td><td>99.1±0.5</td><td>73.8±0.2</td><td>98.3±0.5</td><td>26.0±0.2</td><td>95.8±0.6</td><td>22.5±0.3</td><td>89.4±0.6</td><td>61.0±0.3</td><td>86.1±0.7</td><td>17.9±0.2</td></tr><tr><td>AUTOTUSRoBERTa-Large</td><td>99.4±0.4</td><td>73.9±0.2</td><td>98.9±0.3</td><td>26.7±0.2</td><td>96.2±0.5</td><td>22.3±0.2</td><td>90.0±0.5</td><td>60.5±0.3</td><td>86.9±0.6</td><td>18.1±0.3</td></tr></table>",
+        "bbox": [
+            121,
+            80,
+            880,
+            281
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: MAP@Y and R@Y comparisons (%). Results of AUTOTUS are averaged over five runs.  $\\dagger$ means we replace the base encoder from RoBERTa-Base to BERT-Base, and all models are based on BERT-Base for a fair comparison. Note that perfect R@Y is not possible when the ground truth contains less than Y searched tables.",
+        "bbox": [
+            112,
+            288,
+            882,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "self-supervised pre-training and achieves promising results in various table related NLP tasks. We adopt TaBERT (Yin et al., 2020), TABBIE (Iida et al., 2021), TUTA (Wang et al., 2021), FORTAP (Cheng et al., 2022), and TableFormer (Yang et al., 2022) as baseline encoders to calculate the column unionability score and adopt our Table Scorer module to obtain the table unionability score. Although, there are more baseline models in this category, we just select representative and SOTA ones. The details of baselines are introduced in Appendix B.",
+        "bbox": [
+            112,
+            344,
+            489,
+            521
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            539,
+            324,
+            556
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Overall Performance. Table 2 shows the average results on five table union search benchmarks. The proposed AUTOUS consistently outperforms all baselines in MAP and R performance by leveraging column relational representation learning (with student's T test  $p < 0.05$ ). More specifically, for five benchmarks, compared to the SOTA baseline, AUTOUS on average achieves  $5.1\\%$  higher MAP and  $3.1\\%$  higher R across various benchmarks. Compared with BERT-Base, all table pre-trained and table union search baselines could gain MAP and R performance improvements from better learned tabular representations. The table union search baselines can achieve better results than the table pre-trained baselines by optimizing the column representations specific to the table union search task. AUTOUS exploit the column relational representations and are able to beat all the table union search baselines. In addition, we study the performance changes brought about by using larger base encoders BERT-Large (Devlin et al., 2019) and RoBERTa-Large (Liu et al., 2019).",
+        "bbox": [
+            112,
+            565,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0499b61ea36c5f2efbfdfd3e7170cbf2cd14032c9fd6e44ce96056945e46f517.jpg",
+        "image_caption": [
+            "(a)  $P@K$  on SANTOSN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            341,
+            699,
+            450
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/84830044b7884a84f423356c93293c564612bb0c77016960c91b2789c9f7b0c3.jpg",
+        "image_caption": [
+            "(b)  $R@K$  on SANTOSN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            341,
+            880,
+            451
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/55eca785b1b56006f99c2d36356e329f33382d97f16379c93efa29292cd4d785.jpg",
+        "image_caption": [
+            "(c)  $P@K$  on TUS Small"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            462,
+            692,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8c9a89fea0b3df1cdb4badc7095b352041011c75bfa6556a19a4a31e1cd09ff.jpg",
+        "image_caption": [
+            "(d)  $R@K$  on TUS Small"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            463,
+            880,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/94ed419558c6f3d98114b4c62557686568eade91298fbd229872bc21699a48b1.jpg",
+        "image_caption": [
+            "(e)  $P@K$  on TUS LargeN",
+            "Figure 3:  $\\mathbb{P}@\\mathcal{Y}$  and  $\\mathbb{R}@\\mathcal{Y}$  results on different datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            581,
+            692,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bed15055a2cfe8be30caf286d0486fa7a211596e158b96136bc416126b186a4b.jpg",
+        "image_caption": [
+            "(f)  $R@K$  on TUS LargeN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            581,
+            880,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From Figure 2, it can be concluded that leveraging a large-scale pre-trained language model with more parameters can lead to better MAP and R, BERT-Large and RoBERTa-Large obtain an average boost of  $1.1\\%$  and  $1.4\\%$ , respectively.",
+        "bbox": [
+            507,
+            757,
+            882,
+            835
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Another interesting finding is that, although all the models have decreased performance on the datasets with table noise, the average performance improvement of AUTOTUS compared to SOTA is consistent, and even increased to  $6.6\\%$  on MAP",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3792",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4ce794ff664341b5f89b03f1575dab91fb6d4c58cca9241404feed09b9016e15.jpg",
+        "image_caption": [
+            "(a) AutoTUS w/o Column Pair Classification"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            80,
+            314,
+            178
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/17cdd72e1216f58ac79e11d54b74163a72beaf9a7886f26e323beef0bbfb8487.jpg",
+        "image_caption": [
+            "Figure 4: Visualizing column relational representations after t-SNE dimension reduction on TUS large $_N$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            327,
+            80,
+            505,
+            178
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d1d1ed2a151aa00d423a9d5e1d7618d325b3008b2dbb787906ab8666f09cbdb.jpg",
+        "image_caption": [
+            "(b) AutoTUS w/o Automatic Clustering (c)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            80,
+            694,
+            178
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c255849864abb4d383a01f4e6235217e9c3b634527ae4f5d83b8ca9b8b570132.jpg",
+        "image_caption": [
+            "AutoTUS w/o Real-world Noise Generation",
+            "(d) AutoTUS"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            705,
+            80,
+            875,
+            178
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and  $3.8\\%$  on R. We attribute the improvement of AUTCUS to the noisy, weak column relational representations from the large pre-trained language models are exploited and refined: we bootstrap the representations obtained from noisy tabular data via self-supervised training schema. More specifically, we give the comparison of  $\\mathrm{P}@\\mathrm{Y}$  and  $\\mathrm{R}@\\mathrm{Y}$  as  $Y$  changes in Figure 3. We find that AUTCUS consistently outperforms all baseline models on both metrics  $\\mathrm{P}@\\mathrm{Y}$  and  $\\mathrm{R}@\\mathrm{Y}$  as  $Y$  changes.",
+        "bbox": [
+            112,
+            229,
+            489,
+            391
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Ablation Study. We conduct ablation studies to show the effectiveness of different modules of AUTOTUS. A number of variants of AUTOTUS are considered: AUTOTUS w/o Table Noise Generation removes added table noise in the training data; AUTOTUS w/o Column Relation Encoder Updater via Pseudo Label Classification is AUTOTUS without Column Relation Encoder Updater via Pseudo Label Classification and only uses the Contextualized Column Relation Encoder for Column Relation Encoder Updater via Adaptive Clustering; AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering replaces the suggested soft-assignment clustering techniques with  $k$ -means as a hard-assignment alternative.",
+        "bbox": [
+            112,
+            403,
+            489,
+            643
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The general conclusion from ablation study results shown in Table 2 is that the all three modules contribute positively to improve the performance. More specifically, without table noise in training data, AUTOTUS w/o Table Noise Generation gives  $2.3\\%$  less MAP and R averaged on all datasets, especially the noisy  $\\mathrm{SANTOS}_N$  and TUS large, the drop reaches  $3.5\\%$ . If we stop exploiting self-supervised signals for column relational features learning, AUTOTUS w/o Column Relation Encoder Updater via Pseudo Label Classification brings  $12.8\\%$  less MAP and R averaged on all datasets. This huge performance drop fully demonstrates the importance of learning and refining column-pair representations. Column Relation Encoder Updater via Adaptive Clustering gives  $3.0\\%$  MAP and R boost in average when comparing",
+        "bbox": [
+            112,
+            645,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "with the hard-assignment alternative (AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering).",
+        "bbox": [
+            507,
+            229,
+            884,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Visualizing Contextualized Column Relational Embeddings. To intuitively show how clustering-enhanced self-supervised training can exploit self-supervised signals to obtain better contextualized column relational representations, we visualize the dimensionally reduced column relational representation space  $\\mathbb{R}^{2\\cdot h_R}$  using t-SNE (Maaten and Hinton, 2008). We randomly choose 10 base tables from TUS large $_N$  dataset and sample all column relations that belong to the corresponding base table. We show the visualization results in Figure 4 with each column relation being colored according to their ground-truth.",
+        "bbox": [
+            507,
+            291,
+            884,
+            500
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "From Figure 4, we can see that AUTOTUS w/o Column Relation Encoder Updater via Pseudo Label Classification can assign meaningful semantics to column relations from different base tables, but these unrefined features cannot be tailored for the table union search task. When Column Relation Encoder Updater via Adaptive Clustering is not applied and only  $k$ -means are used, AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering performs a hard assignment of features, and the clustering results appear messy since there are no cluster centroids with high confidence. AUTOTUS w/o Table Noise Generation demonstrates that the clustered features is not tolerate to the noise samples, resulting in unclear cluster boundaries. AUTOTUS shows denser and well-separated clusters, which verifies its powerful column relational representation learning ability.",
+        "bbox": [
+            507,
+            502,
+            884,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Parameter Analysis: when  $\\mathbf{K}$  is unknown. The Column Relation Encoder Updater via Adaptive Clustering provide the flexibility to explore column relational features without knowing any prior information about the number of clusters. This property is attractive when the number of target clusters is agnostic. We vary  $K$  from 10 to 200 and report the",
+        "bbox": [
+            507,
+            806,
+            882,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3793",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8fe6fcada3e668877d0377794b95a349dfa73ad15b264734605fae713da175f.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "AutoTUS (TUS Small) AutoTUS w/o Automatic Clustering (TUS Small)",
+            "AutoTUS (TUS Large) AutoTUS w/o Automatic Clustering (TUS Large)"
+        ],
+        "bbox": [
+            115,
+            108,
+            485,
+            277
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/ca2422e3adb3cb13a1cbb77871d467df8c5912164ccb099d858f505146d115d6.jpg",
+        "table_caption": [
+            "Figure 5: MAP@60 Score with different  $K$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"3\">Table Noise</td><td colspan=\"2\">SANTOSN</td><td colspan=\"2\">TUS largeN</td></tr><tr><td>Column Name</td><td colspan=\"2\">Column Data</td><td>MAP@10</td><td>R@10</td><td>MAP@60</td><td>R@60</td></tr><tr><td>X</td><td colspan=\"2\">X</td><td>83.0</td><td>56.4</td><td>80.7</td><td>14.9</td></tr><tr><td>X</td><td colspan=\"2\">✓</td><td>85.5</td><td>57.6</td><td>82.1</td><td>15.8</td></tr><tr><td>✓</td><td colspan=\"2\">X</td><td>86.9</td><td>59.2</td><td>83.8</td><td>16.4</td></tr><tr><td>✓ or and</td><td colspan=\"2\">✓</td><td>87.6</td><td>59.7</td><td>84.9</td><td>17.3</td></tr><tr><td></td><td></td><td>88.0</td><td>60.1</td><td>84.7</td><td>17.1</td><td></td></tr></table>",
+        "bbox": [
+            117,
+            306,
+            487,
+            401
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Table Noise Analysis.",
+        "bbox": [
+            194,
+            410,
+            405,
+            425
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "MAP@60 score in Figure 5, the best result is obtained when  $K = 25$  for TUS Small and  $K = 50$  for TUS Large, slightly larger than the number of base tables in the two datasets, indicating that AUTOTUS actually exploits the number of target clusters as useful prior knowledge. Thanks to the self-supervised training for column relational features exploitation and the flexibility brought by soft-assignment clustering, when we vary  $K$  from 10 to 200, AUTOTUS gets a more stable MAP@60 score than AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering.",
+        "bbox": [
+            112,
+            437,
+            489,
+            630
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Impact of Table Noise. We investigate the effect of table noise in column value and column name. As shown in Section 3.2, we randomly select a corresponding noise from two types of noises. From Table 3, we observe that adding table noise to both column value and column name help to improve the performance, whereas changing column name has a slightly larger effect (2.8% vs. 1.5%). This improvements may related to the fact that column name summarizes the entire column semantically, so adding noise to the column names can make the model more robust to table noise and thus obtain more performance gains. However, adding more noise did not consistently improve performance, suggesting that the model needs to find a trade-off between noisy data perturbation and increased robustness to noise.",
+        "bbox": [
+            112,
+            645,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            665,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.1 Table Union Search",
+        "text_level": 1,
+        "bbox": [
+            507,
+            111,
+            710,
+            124
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In dataset discovery, it is crucial to find related tables in data lakes. The table union search task has recently received considerable attentions (Ling et al., 2013; Lehmberg and Bizer, 2017; Khatiwada and Fan, 2023). The initial attempt is made by Nargesian et al. (2018). The  $\\mathrm{D}^{3}\\mathrm{L}$  (Bogatu et al., 2020) categories columns into groups by column features. Starmie (Fan et al., 2022) obtains the column features by utilizing a contextualized pretrained language models. SANTOS (Khatiwada and Fan, 2023) leverages a knowledge base to discover the unionable relations between two tables. More recently, a number of table pre-training models have been developed to obtain the contextualized table representation (Yin et al., 2020; Gong et al., 2020; Dong et al., 2022). However, these methods are not tailored to the table union search task.",
+        "bbox": [
+            505,
+            133,
+            885,
+            407
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.2 Self-supervised Learning in NLP",
+        "text_level": 1,
+        "bbox": [
+            507,
+            420,
+            815,
+            436
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Self-supervised learning (SSL) is a method for building models where the output labels are already included in the input data, eliminating the need for additional labeled data (Liu et al., 2021; Hu et al., 2021a,b; Liu et al., 2022d,c). SSL has been widely used in NLP domains such as sentence generation (West et al., 2019; Yan et al., 2021), document processing (You et al., 2021; Ginzburg et al., 2021), natural language inference (Li et al., 2022, 2023), and text reasoning (Klein and Nabi, 2020; Fu et al., 2020; Chen et al., 2022). BERT (Devlin et al., 2019) is one of the most eminent SSL methods which exploit self-supervisions from corpus with next sentence prediction and masked language modeling tasks. In our work, we adopt SSL method to exploit and refine self-supervised signals from tabular data.",
+        "bbox": [
+            505,
+            442,
+            885,
+            715
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            731,
+            650,
+            746
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a multi-stage self-supervised table union search framework, which represents column relation as column relational representation and learn this representation in a multi-stage manner that can better describe column relation for unionability prediction. In additional, a table noise generator is proposed to improve the robustness of our approach against the table noise. The conducted extensive experiments demonstrate the effectiveness of the proposed approach.",
+        "bbox": [
+            505,
+            758,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3794",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            250,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We would like to claim our limitations from two perspectives: technical-wise and application-wise.",
+        "bbox": [
+            112,
+            110,
+            485,
+            141
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Technical-wise: We currently only experiment with BERT-Base, BERT-Large, and RoBERTa-Large as the basic encoders. For larger language models, due to limited resources, we have not implemented them.",
+        "bbox": [
+            112,
+            142,
+            487,
+            221
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Application-wise: The experimental data comes from the Open Data repository released by governments of various countries. Although many domains are covered, some domain-specific data, such as biomedical, have not been considered. Furthermore, our tabular data are all from English, open data research in other languages can be considered as a future research direction.",
+        "bbox": [
+            112,
+            223,
+            487,
+            350
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "8 Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            365,
+            309,
+            382
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We thank the reviewers for their valuable comments. Lijie Wen is the corresponding author. Xuming Hu and Lijie Wen were partially supported by the National Key Research and Development Program of China (No. 2019YFB1704003), the National Nature Science Foundation of China (No. 62021002), Tsinghua BNrist and Beijing Key Laboratory of Industrial Bigdata System and Application. Philip S. Yu was partially supported by the NSF under grants III-1763325, III-1909323, III-2106758, SaTC-1930941.",
+        "bbox": [
+            112,
+            391,
+            489,
+            565
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            596,
+            213,
+            611
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Rami Aly, Zhijiang Guo, Michael Sejr Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. Feverous: Fact extraction and verification over unstructured and structured information. In Proc. of NeurIPS: Datasets and Benchmarks Track.",
+            "Alex Bogatu, Alvaro AA Fernandes, Norman W Paton, and Nikolaos Konstantinou. 2020. Dataset discovery in data lakes. In Proc. of ICDE, pages 709-720. IEEE.",
+            "Dan Brickley, Matthew Burgess, and Natasha Noy. 2019. Google dataset search: Building a search engine for datasets in an open web ecosystem. In Proc. of WWW, pages 1365-1375.",
+            "Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020. Hybridqa: A dataset of multi-hop question answering over tabular and textual data. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1026-1036."
+        ],
+        "bbox": [
+            115,
+            619,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yubo Chen, Yunqi Zhang, and Yongfeng Huang. 2022. Learning reasoning patterns for relational triple extraction with mutual generation of text and graph. In *Findings of ACL*, pages 1638-1647.",
+            "Zhoujun Cheng, Haoyu Dong, Ran Jia, Pengfei Wu, Shi Han, Fan Cheng, and Dongmei Zhang. 2022. Fortap: Using formulas for numerical-reasoning-aware table pretraining. In Proc. of ACL, pages 1150-1166.",
+            "Nadiia Chepurko, Ryan Marcus, Emanuel Zgraggen, Raul Castro Fernandez, Tim Kraska, and David Karger. 2020. Arda: automatic relational data augmentation for machine learning. Proc. of VLDB, 13(9):1373-1387.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. of NAACL-HLT, pages 4171-4186.",
+            "Haoyu Dong, Zhoujun Cheng, Xinyi He, Mengyu Zhou, Anda Zhou, Fan Zhou, Ao Liu, Shi Han, and Dongmei Zhang. 2022. Table pretraining: A survey on model architectures, pretraining objectives, and downstream tasks. arXiv preprint arXiv:2201.09745.",
+            "Grace Fan, Jin Wang, Yuliang Li, Dan Zhang, and Renee Miller. 2022. Semantics-aware dataset discovery from data lakes with contextualized column-based representation learning. arXiv preprint arXiv:2210.01922.",
+            "Tsu-Jui Fu, Xin Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. 2020. Sscr: Iterative language-based image editing via self-supervised counterfactual reasoning. In Proc. of EMNLP, pages 4413-4422.",
+            "Sainyam Galhotra and Udayan Khurana. 2020. Semantic search over structured data. In Proc. of CIKM, pages 3381-3384.",
+            "Dvir Ginzburg, Itzik Malkiel, Oren Barkan, Avi Caci-ularu, and Noam Koenigstein. 2021. Self-supervised document similarity ranking via contextualized language models and hierarchical inference. In *Findings of ACL-IJCNLP*, pages 3088-3098.",
+            "Heng Gong, Yawei Sun, Xiaocheng Feng, Bing Qin, Wei Bi, Xiaojiang Liu, and Ting Liu. 2020. Tablegpt: Few-shot table-to-text generation with table structure reconstruction and content matching. In Proc. of COLING, pages 1978-1988.",
+            "Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178-206.",
+            "Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Mueller, Francesco Piccinno, and Julian Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. In Proc. of ACL, pages 4320-4333."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3795",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xuming Hu, Lijie Wen, Yusong Xu, Chenwei Zhang, and Philip S. Yu. 2020. Selfore: Self-supervised relational feature learning for open relation extraction. In Proc. of EMNLP, pages 3673-3682.",
+            "Xuming Hu, Chenwei Zhang, Fukun Ma, Chenyao Liu, Lijie Wen, and Philip S. Yu. 2021a. Semi-supervised relation extraction via incremental meta self-training. In Findings of EMNLP, pages 487-496.",
+            "Xuming Hu, Chenwei Zhang, Yawen Yang, Xiaohe Li, Li Lin, Lijie Wen, and Philip S. Yu. 2021b. Gradient imitation reinforcement learning for low resource relation extraction. In Proc. of EMNLP, pages 2737-2746.",
+            "Hiroshi Iida, Dung Thai, Varun Manjunatha, and Mohit Iyyer. 2021. Tabbie: Pretrained representations of tabular data. In Proc. of NAACL-HLT, pages 3446-3456.",
+            "Aamod Khatiwada and Grace Fan. 2023. Santos: Relationship-based semantic table union search. In Proc. of SIGMOD.",
+            "Tassilo Klein and Moin Nabi. 2020. Contrastive self-supervised learning for commonsense reasoning. In Proc. of ACL, pages 7517-7523.",
+            "Christos Koutras, George Siachamis, Andrea Ionescu, Kyriakos Psarakis, Jerry Brons, Marios Fragkoulis, Christoph Lofi, Angela Bonifati, and Asterios Katsifodimos. 2021. Valentine: Evaluating matching techniques for dataset discovery. In Proc. of ICDE, pages 468-479. IEEE.",
+            "Oliver Lehmberg and Christian Bizer. 2017. Stitching web tables for improving matching quality. Proc. of VLDB, 10(11):1502-1513.",
+            "Shu'ang Li, Xuming Hu, Li Lin, Aiwei Liu, Lijie Wen, and Philip S. Yu. 2023. A multi-level supervised contrastive learning framework for low-resource natural language inference. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:1771-1783.",
+            "Shu'ang Li, Xuming Hu, Li Lin, and Lijie Wen. 2022. Pair-level supervised contrastive learning for natural language inference. arXiv preprint arXiv:2201.10927.",
+            "Xiao Ling, Alon Y Halevy, Fei Wu, and Cong Yu. 2013. Synthesizing union tables from the web. In Proc. of IJCAI.",
+            "Aiwei Liu, Xuming Hu, Li Lin, and Lijie Wen. 2022a. Semantic enhanced text-to-sql parsing via iteratively learning schema linking graph. In Proc. of KDD, pages 1021-1030.",
+            "Aiwei Liu, Xuming Hu, Lijie Wen, and Philip S Yu. 2023. A comprehensive evaluation of chatgpt's zero-shot text-to-sql capability. arXiv preprint arXiv:2303.13547."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            916
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ruixue Liu, Shaozu Yuan, Aijun Dai, Lei Shen, Tiangang Zhu, Meng Chen, and Xiaodong He. 2022b. Few-shot table understanding: A benchmark dataset and pre-training baseline. In Proc. of COLING, pages 3741-3752.",
+            "Shuliang Liu, Xuming Hu, Chenwei Zhang, Shu'ang Li, Lijie Wen, and Philip S. Yu. 2022c. Hiure: Hierarchical exemplar contrastive learning for unsupervised relation extraction. In Proc. of NAACL-HLT, pages 5970-5980.",
+            "Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. 2021. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering.",
+            "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.",
+            "Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip Yu. 2022d. Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering.",
+            "Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(Nov):2579-2605.",
+            "Fatemeh Nargesian, Erkang Zhu, Ken Q Pu, and Renee J Miller. 2018. Table union search on open data. Proc. of VLDB, 11(7):813-825.",
+            "Aécio Santos, Aline Bessa, Christopher Musco, and Juliana Freire. 2022. A sketch-based index for correlated dataset search. In Proc. of ICDE, pages 2928-2941. IEEE.",
+            "Livio Baldini Soares, Nicholas Fitzgerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In Proc. of ACL, pages 2895-2905.",
+            "Zhiruo Wang, Haoyu Dong, Ran Jia, Jia Li, Zhiyi Fu, Shi Han, and Dongmei Zhang. 2021. Tuta: tree-based transformers for generally structured table pretraining. In Proc. of SIGKDD, pages 1780-1790.",
+            "Peter West, Ari Holtzman, Jan Buys, and Yejin Choi. 2019. Bottlesum: Unsupervised and self-supervised sentence summarization using the information bottleneck principle. In Proc. of EMNLP-IJCNLP, pages 3752-3761.",
+            "Yihai Xi, Ning Wang, Shuang Hao, Yiyi Zhang, and Xinyu Chen. 2023. Popularity sensitive and domain-aware summarization for web tables. Information Sciences, 621:729-748.",
+            "Junyuan Xie, Ross Girshick, and Ali Farhadi. 2016. Unsupervised deep embedding for clustering analysis. In Proc. of ICML, pages 478-487."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3796",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. 2021. Consent: A contrastive framework for self-supervised sentence representation transfer. In Proc. of ACL-IJCNLP, pages 5065-5075.",
+            "Jingfeng Yang, Aditya Gupta, Shyam Upadhyay, Luheng He, Rahul Goel, and Shachi Paul. 2022. Tableformer: Robust transformer modeling for table-text encoding. In Proc. of ACL, pages 528-537.",
+            "Pengcheng Yin, Graham Neubig, Wen-tau Yih, and Sebastian Riedel. 2020. Tabert: Pretraining for joint understanding of textual and tabular data. In Proc. of ACL, pages 8413-8426.",
+            "Chenyu You, Nuo Chen, and Yuexian Zou. 2021. Self-supervised contrastive cross-modality representation learning for spoken question answering. In *Findings of EMNLP*, pages 28-39.",
+            "Dan Zhang, Madelon Hulsebos, Yoshihiko Suhara, Cagatay Demiralp, Jinfeng Li, and Wang-Chiew Tan. 2020. Sato: contextual semantic type detection in tables. Proceedings of the VLDB Endowment, 13(12):1835-1848.",
+            "Shuo Zhang and Krisztian Balog. 2017. Entitables: Smart assistance for entity-focused tables. In Proc. of SIGIR, pages 255-264.",
+            "Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. 2021. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. In Proc. of ACL-IJCNLP, pages 3277-3287."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            552
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "3797",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Dataset Construction Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            401,
+            98
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "For test data, the SANTOS Small benchmark (Khatiwada and Fan, 2023) consists of 550 data lake tables, which are generated from 296 open datasets from Canada, the United Kingdom, the United States, and Australia. The SANTOS Small also has 50 query tables. There are two benchmarks accessible from Nargesian et al. (2018): TUS Small and TUS Large. The TUS Small consists of 1,530 data lake tables that are generated from 10 base table of Canada open data. The TUS Small also has 150 query tables. The TUS Large consists of 5,043 data lake tables that are generated from 32 base table of Canada open data. The TUS Large also has 100 query tables. The SANTOS $^{4}$  and TUS $^{5}$  benchmarks and their unionable table ground truth are publicly available. For training data, we randomly generated the same number of unlabeled training tables as the labeled tables from base tables of Open Data. Note that the public code for generating tables is publicly available $^{6}$  (Koutras et al., 2021; Khatiwada and Fan, 2023). Since table union search task is an unsupervised task, no dev data is obtained.",
+        "bbox": [
+            115,
+            108,
+            489,
+            476
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B The Introduction of Baseline Models",
+        "text_level": 1,
+        "bbox": [
+            114,
+            489,
+            465,
+            504
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We compare AUTOTUS with two categories of baseline models. The first category of models leverage unlabeled tables for self-supervised pretraining and achieve promising results in the table understanding tasks as baseline encoders to calculate the column unionability score:",
+        "bbox": [
+            112,
+            514,
+            489,
+            609
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) TaBERT (Yin et al., 2020) is a pretrained language model that simultaneously learns representations for (semi-)structured tables and natural language phrases. 26 million tables and their English contexts make up the vast corpus on which TaBERT was trained.",
+            "(2) TABBIE (Iida et al., 2021) develops a straightforward pretraining target (corrupt cell identification) that only learns from tabular data and achieves the state-of-the-art on the table-based tasks. TABBIE offers embeddings of all table substructures (cells, rows, and columns), unlike competing techniques, and it also takes far less computing power to train."
+        ],
+        "bbox": [
+            112,
+            611,
+            489,
+            835
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(3) TUTA (Wang et al., 2021) is a unified pretraining architecture for comprehending typically arranged tables. TUTA improves transformers with three structure-aware techniques after realizing that understanding a table necessitates spatial, hierarchical, and semantic information.",
+            "(4) FORTAP (Cheng et al., 2022) explores to leverage the spreadsheet formulas for table pretraining. FORTAP adopts two self-supervised pretraining objectives, which are derived from formulas, numerical reference prediction and numerical calculation prediction.",
+            "(5) TableFormer (Yang et al., 2022) is a structurally conscious table-text encoding architecture in which learnable attention biases are used to fully include tabular structural biases."
+        ],
+        "bbox": [
+            507,
+            84,
+            884,
+            340
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Of course, there are more baseline models in this category, we just select representative and SOTA models. The second category of models adopt various column representations to calculate the column unionability score:",
+        "bbox": [
+            507,
+            342,
+            882,
+            420
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(6)  $\\mathbf{D}^3\\mathbf{L}$  (Bogatu et al., 2020) creates hash-based indexes using the features of the items in a dataset and maps those features to a uniform distance space.",
+            "(7) SATO (Zhang et al., 2020) is a hybrid machine learning model that uses both the context of the table and the values of the columns to automatically identify the semantic categories of columns in tables.",
+            "(8) Starmie (Fan et al., 2022) obtains the semantic information included inside tables by utilizing a contrastive multi-column pre-training technique.",
+            "(9) SANTOS (Khatiwada and Fan, 2023) suggests a definition of unionability that takes connections between columns and their semantics into principled consideration."
+        ],
+        "bbox": [
+            507,
+            422,
+            882,
+            678
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_footnote",
+        "text": "4https://github.com/   \nnortheastern-datalab/santos   \n5https://github.com/RJMillerLab/   \ntable-union-search-benchmark   \n6https://delftdata.github.io/   \nvalentine/",
+        "bbox": [
+            112,
+            843,
+            426,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "3798",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? Section 7",
+            "A2. Did you discuss any potential risks of your work? Section 7",
+            "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            287
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            299,
+            487,
+            316
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section 3 and Section 4",
+        "bbox": [
+            132,
+            321,
+            310,
+            335
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? Section 3 and Section 4",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3 and Section 4",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3 and Section 4",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Section 3, Section 4, and Appendix A",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4, and Appendix A"
+        ],
+        "bbox": [
+            129,
+            346,
+            880,
+            753
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section 4",
+        "bbox": [
+            132,
+            787,
+            206,
+            800
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4",
+        "bbox": [
+            129,
+            812,
+            880,
+            859
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            866,
+            877,
+            889
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "3799",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4 and Appendix A",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank."
+        ],
+        "bbox": [
+            127,
+            83,
+            878,
+            282
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.",
+        "bbox": [
+            112,
+            293,
+            877,
+            330
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response."
+        ],
+        "bbox": [
+            127,
+            340,
+            878,
+            640
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "3800",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2023/Automatic Table Union Search with Tabular Representation Learning/7244320a-8cb5-4ef7-87f0-c4d2dc4ad4ba_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94618e370b9c3a6f229fc94ff6519d3a0863f6d2af6675afb72d82cab8f03de9
+size 1167330

2023/Automatic Table Union Search with Tabular Representation Learning/full.md

@@ -0,0 +1,391 @@
+# Automatic Table Union Search with Tabular Representation Learning
+
+Xuming Hu<sup>1*</sup>, Shen Wang<sup>2</sup>, Xiao Qin<sup>2</sup>, Chuan Lei<sup>2</sup>, Zhengyuan Shen<sup>2</sup>, Christos Faloutsos<sup>2</sup>, Asterios Katsifodimos<sup>2</sup>, George Karypis<sup>2</sup>, Lijie Wen<sup>1</sup>, Philip S. Yu<sup>1,3</sup>
+
+$^{1}$ Tsinghua University, $^{2}$ Amazon Web Services, $^{3}$ University of Illinois at Chicago hxml9@mails.tsinghua.edu.cn, wenlj@tsinghua.edu.cn, psyu@uic.edu  $\{$  shenwa, drxqin, chuanlei, donshen\} @amazon.com  $\{$  faloutso, akatsifo, gkarypis\} @amazon.com
+
+# Abstract
+
+Given a data lake of tabular data as well as a query table, how can we retrieve all the tables in the data lake that can be unioned with the query table? Table union search constitutes an essential task in data discovery and preparation as it enables data scientists to navigate massive open data repositories. Existing methods identify uniability based on column representations (word surface forms or token embeddings) and column relation represented by column representation similarity. However, the semantic similarity obtained between column representations is often insufficient to reveal latent relational features to describe the column relation between pair of columns and not robust to the table noise. To address these issues, in this paper, we propose a multi-stage self-supervised table union search framework called AUTOTUS, which represents column relation as a vector- column relational representation and learn column relational representation in a multi-stage manner that can better describe column relation for table unionability prediction. In particular, the large language model powered contextualized column relation encoder is updated by adaptive clustering and pseudo label classification iteratively so that the better column relational representation can be learned. Moreover, to improve the robustness of the model against table noises, we propose table noise generator to add table noise to the training table data. Experiments on real-world datasets and synthetic test set augmented with table noise show that AUTOTUS achieves  $5.2\%$  performance gain over the SOTA baseline.
+
+# 1 Introduction
+
+The growing availability of tabular data from academia and industry brings new opportunities for economic growth and social benefit, which has also attracted considerable attention in the natural language processing (NLP) community. A number of
+
+Query Table  
+
+<table><tr><td>Name</td><td>BP</td><td>Team</td><td>Jersey Number</td><td>CLG</td></tr><tr><td>Kevon Looney</td><td>Milwaukee</td><td>Warriors</td><td>5</td><td>UCUCLA</td></tr><tr><td>Anthony Davis</td><td>Chicago</td><td>Lakers</td><td>3</td><td>KentKentucky</td></tr><tr><td>Devin Booker</td><td>Moss Point</td><td>Suns</td><td>1</td><td>KentKentucky</td></tr></table>
+
+Data Lake Table (Unionable)  
+
+<table><tr><td>Player</td><td>Birth Place</td><td>TM</td><td>POS</td><td>Draft</td></tr><tr><td>Andrew Wiggins</td><td>Toronto</td><td>Golden State Warriors</td><td>Shootung Guard</td><td>R1-1</td></tr><tr><td>Shai Gilgeous-Alexander</td><td>Hamilton</td><td>Oklahoma City Thunder</td><td>Point Gusrd</td><td>R1-11</td></tr><tr><td>Dillon Brooks</td><td>Mississauga</td><td>Memphis Grizzlies</td><td>Small Forward</td><td>R2-45</td></tr></table>
+
+![](images/6d62e8a5f7b52a5885cf46228c272af899dd68e4ea25ac4bc88648352c821042.jpg)  
+Figure 1: The table union search example on Open Data.
+
+work has been done, such as Table QA (Chen et al., 2020; Zhu et al., 2021), Table fact verification (Aly et al., 2021; Guo et al., 2022), Table summarization (Xi et al., 2023), Table understanding (Yang et al., 2022; Liu et al., 2022a, 2023), etc. More recently, table union search (Nargesian et al., 2018; Bogatu et al., 2020; Khatiwada and Fan, 2023) is proposed to facilitate tabular dataset discovery applications (Brickley et al., 2019; Galhotra and Khurana, 2020; Santos et al., 2022), which can benefit the tabular data integration and analysis. The table union search aims to find all tables in a data lake that have the columns from the same domain as the query table. With the help of the table union search, the understanding of the tabular data is largely improved and it can potentially benefit other tabular-focused NLP downstream tasks. Therefore, table union search is a non-trivial NLP research problem.
+
+Recent literature shows that column representation methods are useful for table union search. The basic idea is to map the column to a latent vector space (e.g. column representations (Bogatu et al., 2020; Chepurko et al., 2020) or token embeddings (Zhang and Balog, 2017; Herzig et al., 2020; Fan et al., 2022; Yang et al., 2022)) and then compute the table unionability score based on the column relation obtained from the similarity scores of col
+
+unn pairs between query table and target tables. However, these methods only consider the column representation and their similarity as the column relation to compute the table unionability score, which is not enough to describe the relationship between two columns and may cause incorrect prediction. Figure 1 shows an example of the table union search with a query table and a unionable table. By looking at the second columns (in red) of two tables, the computed column similarity score (a scalar) from column embedding is low due to the dissimilar token embeddings. In this case, the column relation is not intuitive, taking only the similarity between column embeddings to describe the column relation is not enough to predict the table unionability accurately. However, if we combine each column representation and model them as an embedding (a vector), the new representation (column relational representation) is more informative and can better describe the unintuitive column relation to make the table unionability prediction more accurately. Therefore, a better approach is needed, which can make the table unionability prediction via the column relational representation.
+
+On the other hand, real-world table noises widely exist (Koutras et al., 2021; Liu et al., 2022b). Taking column "CLG" in the query table and column "POS" in the data lake table as examples, both columns have abbreviated column names, tokens are repeated in column values (e.g., "UCLA" becomes "UCUCLA") and proximal characters in keyboard are replaced (e.g., "Shooting Guard" becomes 'Shootung Guard'). Computing column representations or column relational representations based on token embeddings with tabular noise introduces unavoidable wrong predictions. Therefore, it requires the table union search model to be robust to the table noise during optimization.
+
+To address above issues, we propose a multi-stage self-supervised table union search framework called AutotUS, which represents column relation as a vector- column relational representation and learn column relational representation in a multi-stage manner that can better describe column relation for unionability prediction. In particular, the large language model powered contextualized column relation encoder is updated by adaptive clustering and pseudo label classification iteratively so that the better column relational representation can be learned. Moreover, to improve the robustness of the model against table noises, we pro
+
+pose table noise generator to add table noise to the training table data. Extensive experiments are conducted on three real-world datasets and show significant performance gains compared to existing state-of-the-art methods and demonstrates the robustness of AUTOTUS against the table noise.
+
+Our contributions are four-folded: 1) We propose a novel framework AUTOTUS that can leverage column relational representation for table union search instead of pairwise column similarity. 2) We propose a multi-stage self-supervised method that can update the large language model powered contextualized column relation encoder in a step by step and iterative way to learn better column relational representation. 3) We propose a table noise generator and add table noise to the training tabular data to improve the robustness of the model against the table noise. 4) We conduct extensive experiment on real-world datasets as well as synthetic test set augmented with table noise and demonstrate significant performance gain of AUTOTUS over the strong baselines and shows the robustness of AUTOTUS against the table noise.
+
+# 2 Problem Definition
+
+Table union search discovers all tables in a set of data lake tables that are工商联able with the query table. Specifically, we focus on the automatically case where no labeled training data is available. We define a set of data lake tables as  $\mathcal{T}$ . For each table  $T\in \mathcal{T}$  which consists of  $m$  columns  $\{t_1,t_2,\dots ,t_m\}$ . There is a column relation  $e_{ij}\in \mathcal{E}$  for each pairs of columns  $t_i,t_j$  that describes their relationship. We employ a contextualized column relational encoder to obtain the column relational representation  $\mathbf{z}_{e_{ij}}\in \mathcal{Z}$  for each column relation  $e_{ij}$ . The table unionability score  $u_{(T_i,T_j)}\in \mathcal{U}$  is obtained by the table scorer that goes over all column relational representations between two tables  $T_{i},T_{j}$ . Similar to previous studies (Nargesian et al., 2018; Bogatu et al., 2020; Khatiwada and Fan, 2023), given a query table  $Q$ , table union search retrieves the top- $y$ 工商联able tables  $\mathcal{Q}\subseteq \mathcal{T}$ , where  $|\mathcal{Q}| = y$  and  $\forall T\in \mathcal{Q}$  and  $T^{\prime}\in \mathcal{T} - \mathcal{Q}$ , we have  $u_{(Q,T)}\geq u_{(Q,T^{\prime})}$ .
+
+# 3 Proposed Method
+
+In this section, we introduce the proposed AU-TOTUS. The framework (as shown in Figure 2) consists of following components: Multi-stage Self-supervised Column Relational Representation
+
+![](images/4e07cc2860c3430a6eff1d42b3093411838d5ab2f9a7b22466d0cfe22a1f08d2.jpg)  
+Figure 2: Overview of AUTOTUS framework.
+
+Learning, Table Noise Generation and Table Unionability Scorer. We first illustrate the motivation for AUTOTUS: The existing column representation methods use cosine similarity to measure the unionability of column pairs. The column encoder in these methods is usually only trained with a binary classification task, i.e. whether two columns are unionable. In AUTOTUS, the table encoder can benefit from an additional task: the identification of hidden semantic types (domains, task knowledge, etc.) from the column pairs. In the absence of explicit type labels, we use adaptive clustering on column pairs and learn those hidden types. The classification step is then used to update the table encoder, so that the embeddings are generated in a way that minimizes the hidden type prediction errors. In short, the clustering and classification modules learn the hidden unionable semantic types, resulting in a better table encoder for the table union search task.
+
+# 3.1 Multi-stage Self-supervised Column Relational Representation Learning
+
+The Multi-stage Self-supervised Column Relational Representation Learning module aims to extract column relation as column relational representation and learn column relational representation in multi-stage manner that can better describe column relation for unionability prediction. As illustrated in Figure 2, this module consists of three components: Contextualized Column Relation Encoder, Column Relation Encoder Updater via Adaptive Clustering, and Column Relation Encoder Updater via Pseudo Label Classification. The Contextualized Column Relation Encoder serializes and tokenizes the column relations as the input and leverages the pretrained BERT (Devlin et al., 2019) to generate contextualized column relation embeddings. The Column Relation Encoder Updater via
+
+Adaptive Clustering optimizes Contextualized Column Relation Encoder through relaxed constraint to obtain better column relational representation and generate the cluster label as the pseudo label. The Column Relation Encoder Updater via Pseudo Label Classification uses the generated pseudo label to further update the Contextualized Column Relation Encoder via more strict constraint to generate improved column relational representation. The above two updates are performed in a bootstrapped loop that iteratively improve the language model to generate better column representations.
+
+Contextualized Column Relation Encoder Contextualized Column Relation Encoder extracts the contextualized column relation embeddings to represent the relational features between two columns. We leverage a large language model, BERT (Devlin et al., 2019), to effectively encode column relations, along with their context tabular information.
+
+The pre-trained large language models (such as BERT) support an input length of up to 512 tokens (Devlin et al., 2019). However, columns in a real-world table may contain thousands of tokens, and randomly discarding tokens will lead to loss of important semantics. Therefore, in this work, we score the importance of each token or cell in a column, and keep cells based on their importance. More specifically, we adopt TF-IDF score to calculate the token importance through its inverse document frequency:  $\log(|\mathcal{Y}|)/\{t|\text{token} \in t, t \in \mathcal{Y}\}$ , where  $t$  is a column and  $|\mathcal{Y}|$  is the number of all data lake columns. Then the cell score is calculated by averaging the TF-IDF of all tokens within the cell. After obtaining the importance score of each column-level cell, the row alignment can ensure the correctness of the table semantics, e.g.: Jimmy Butler's birthplace is aligned to the United States instead of the United Kingdom. Therefore, we cal
+
+culate the average score of cells in each row and sort them in descending order. Next, we select rows until the total token count reaches the input budget and obtain the table:  $T = [t_{1}, t_{2}, \dots, t_{m}]$ .
+
+As shown in Figure 2, we serialize the table by column and follow the marking schema adopted in Soares et al. (2019) to augment  $T$  with two reserved tokens to mark the start and the end of each column. We introduce  $\langle s \rangle$  and  $\langle /s \rangle$  and inject them into  $T$ :
+
+$$
+T = \left[ \langle s \rangle , t _ {1}, \langle / s \rangle , \dots , \langle s \rangle , t _ {m}, \langle / s \rangle \right], \tag {1}
+$$
+
+as the input token sequence for the encoder.
+
+We denote the Contextualized Column Relation Encoder as  $f_{\theta}(T,\langle s\rangle ,\langle \langle s\rangle \rangle)$ . Instead of using the output of [CLS] token from BERT itself which summarizes the whole table-level semantics, we get the contextualize column relation semantics by concatenating the representations  $\langle s\rangle$  of any two columns in different tables or in the same table, and derive a fixed-length representation  $\mathbf{z}_{e_{ij}}\in \mathbb{R}^{2\cdot z_R}$ :
+
+$$
+\mathbf {z} _ {e _ {i j}} = \left[ \mathbf {c} _ {\langle s _ {t _ {i}} \rangle}, \mathbf {c} _ {\langle s _ {t _ {j}} \rangle} \right]. \tag {2}
+$$
+
+Column Relation Encoder Updater via Adaptive Clustering The Column Relation Encoder Updater via Adaptive Clustering encourages each low-confidence relation label distribution to be closer to the high-confidence cluster centroids, improves the confidence of the clusters, and optimizes Contextualized Column Relation Encoder by relax-constraints to generate better column relational representations. In addition, the cluster label generated for each column relation could be treated as self-supervision signal, which serves as the pseudo label for language model update.
+
+More specifically, after obtaining  $N$  column relational representations  $\mathcal{Z} = \{\mathbf{z}_1,\mathbf{z}_2,\dots,\mathbf{z}_N\}$ , Column Relation Encoder Updater via Adaptive Clustering computes  $K$  cluster centroids  $\{\xi_k\in \mathbb{R}^{h_{AC}}\}_{k = 1}^K$ , and softly assigns all  $N$  column relation representations to  $K$  clusters. In practice, we first adopt standard  $k$ -means clustering among  $N$  column relation data points  $\mathbb{R}^{h_{AC}}$  to obtain  $K$  initial cluster centroids  $\{\xi_k\in \mathbb{R}^{h_{AC}}\}_{k = 1}^K$ . Inspired by Maaten and Hinton (2008); Hu et al. (2020), we leverage the Student's  $t$ -distribution as the kernel to measure the similarity of the embedding point  $\mathbf{z}_n$  to each centroid  $\xi_{k}$  as:
+
+$$
+q _ {n k} = \frac {\left(1 + \left\| \mathbf {z} _ {n} - \xi_ {k} \right\| ^ {2} / \alpha\right) ^ {- \frac {\alpha + 1}{2}}}{\sum_ {k ^ {\prime} = 1} ^ {K} \left(1 + \left\| \mathbf {z} _ {n} - \xi_ {k ^ {\prime}} \right\| ^ {2} / \alpha\right) ^ {- \frac {\alpha + 1}{2}}}, \tag {3}
+$$
+
+where  $q_{nk}$  can be viewed as the probability of assigning column relation embedding  $\mathbf{z}_n$  to the cluster  $k$  as the soft assignment and  $\alpha$  denotes the degree of the Student's  $t$ -distribution. We set  $\alpha = 1$  for all experiments.
+
+We normalize each cluster by adopting the frequency as an auxiliary target distribution proposed by Xie et al. (2016) in Equation 4 and iteratively refine each cluster with the help of the auxiliary target distribution:
+
+$$
+p _ {n k} = \frac {q _ {n k} ^ {2} / f _ {k}}{\sum_ {k ^ {\prime}} q _ {n k ^ {\prime}} ^ {2} / f _ {k ^ {\prime}}}, \tag {4}
+$$
+
+where  $f_{k} = \sum_{n = 1}^{N}q_{nk}, k = 1,2,\dots ,K$  is the soft cluster frequency. With the auxiliary target distribution, we could encourage each cluster to learn from its high confidence cluster assignments and simultaneously alleviating the bias caused by imbalanced clusters. We define KL divergence loss between the soft assignments  $q_{n}$  and the auxiliary distribution  $p_n$  to train the Contextualized Column Relation Encoder as follows:
+
+$$
+\mathcal {L} _ {A C} = K L (P | | Q) = \sum_ {n} \sum_ {k} p _ {n k} \log \frac {p _ {n k}}{q _ {n k}}. \tag {5}
+$$
+
+We adopt the label associated with the largest probability as the pseudo label  $\mu_{n}$  for the  $n$ -th column relation:
+
+$$
+\mu_ {n} = \underset {k \in K} {\operatorname {a r g m a x}} p _ {n k}. \tag {6}
+$$
+
+In summary, comparing with traditional clustering methods such as  $k$ -means, Column Relation Encoder Updater via Adaptive Clustering adopts an iterative, soft-assignment learning process which encourages high-confidence assignments and uses high-confidence assignments to improve low confidence ones. Therefore, Column Relation Encoder Updater via Adaptive Clustering is robust enough to different cluster numbers  $K$ , thus solving the situation that we have no prior knowledge of specific domains and the relational distributions in the column relational representations. To mitigate the possible negative impact of choosing a unideal initial centroid, Column Relation Encoder Updater via Adaptive Clustering randomly re-selects a set of  $K$  initial centroids if  $\mathcal{L}_{AC}$  does not decrease after the first epoch.
+
+Column Relation Encoder Updater via Pseudo Label Classification Although the clustering module forces the relational label distribution of
+
+low confidence to be close to the cluster centroids of high confidence to improve the confidence of clustering. However, the clusters may contain some different labels, which affects the clustering purity. Therefore, we utilize the Column Relation Encoder Updater via Pseudo Label Classification module to classify the pseudo labels, and obtain better column relational representations by improving the clustering purity. In practice, the Column Relation Encoder Updater via Adaptive Clustering generates cluster labels  $M = \{\mu_1,\mu_2,\dots,\mu_N\}$  for all column relations as pseudo labels. Column Relation Encoder Updater via Pseudo Label Classification could adopt these pseudo labels as self-supervisions derived from the corpora themselves and use pseudo labels to guide the feature representation learning in Contextualized Column Relation Encoder as well as column relation classifier learning in Column Relation Encoder Updater via Pseudo Label Classification.
+
+The Column Relation Encoder Updater via Pseudo Label Classification module learns to predict the pseudo labels as golden labels. More specifically, we have:
+
+$$
+\mathbf {l} _ {n} = \tau_ {\phi} (f _ {\theta} (T, \langle s \rangle , \langle / s \rangle)), \tag {7}
+$$
+
+where  $\tau_{\phi}$  denotes the column relation classification module with parameters  $\phi$  and  $\mathbf{l}_n$  is the probability distribution for the  $n$ -th sample over  $K$  pseudo labels. To find the best performance parameters  $\theta$  for Contextualized Column Relation Encoder and  $\phi$  for the classifier, we optimize the classification loss:
+
+$$
+\mathcal {L} _ {R C} = \min _ {\theta , \phi} \frac {1}{N} \sum_ {n = 1} ^ {N} l o s s (\mathbf {l} _ {n}, \mathrm {o n e \_ h o t} (\mu_ {n})), (8)
+$$
+
+where loss is the cross entropy loss and one-hot  $(\mu_{n})$  returns a one-hot pseudo label assignment vector.
+
+# 3.2 Table Noise Generator
+
+The table noise generator module aims to improve the robustness of the model against the table noise. We firstly summarizes the noise types that exist in the real-world tabular data, and introduce these table noises to the column relations during the self-supervised training. Specifically, we defined two types of table noise: column value noise an column name noise. Here are some representative column value noises: 1) Replace / insert characters with proximal characters, e.g., computer  $\rightarrow$
+
+computer / cocomputer. 2) Delete / repeat characters, e.g., computer  $\rightarrow$  compuer / computer. 3) Change the numeral display format like scientific notation, e.g.,  $12000\to 1.2\mathrm{e}4.$  Here are some representative column name noises: 1) Prefix column names with table name, e.g., traveldestination  $\rightarrow$  tourism_statistics_traveldestination. 2) Abbreviate column names, e.g., traveldestination  $\rightarrow$  traDest. 3) Drop vowels, e.g., traveldestination  $\rightarrow$  trvl_dstntn. 4) Synonym substitution (if exists), e.g., traveldestination  $\rightarrow$  touristdestination. 5) Acronym, e.g., traveldestination  $\rightarrow$  t_d.
+
+For each table  $T$  in the data lake tables  $\mathcal{T}$ , we randomly apply above noise to the training tabular data to improve the robustness of our model against the table noise.
+
+# 3.3 Table Unionability Scorer
+
+The Table Scorer obtains the table unionability score based on the column relational representations between column relations in the query table and data lake tables, which can provide more sufficient information to clarify whether the two tables come from the same domain and have a union relation. After we obtained the Contextualized Column Relation Encoder with discriminative power to get the column relation embeddings  $\mathbf{z}_{e_{ij}} = (\mathbf{c}_{\langle s_{t_i}\rangle},\mathbf{c}_{\langle s_{t_j}\rangle})$  , we adopt the Column Relation Encoder Updater via Pseudo Label Classification to assign the  $K$  relational labels to all  $N$  column relations  $\mathbf{z}^n = (\mathbf{z}_{e_{ij}})^n$  in the two tables:
+
+$$
+\psi_ {n k} = \underset {k \in K, n \in N} {\operatorname {a r g m a x}} \tau_ {\phi} \left(\mathbf {z} ^ {n k}\right), \tag {9}
+$$
+
+and assign the majority label  $\kappa$  as the predict relational label for the query table and data lake table. We count the number of majority labels  $\kappa$  in the all  $N$  column relational representations between two tables as  $M$ , so the table unionability score is  $M / N$ . This score no longer depends on the calculation of cosine similarity between two column features to obtain the column unionability score, but obtains table unionability score through the column unionability scores. The higher the score, the greater the probability of these two tables being unjoinable. Instead of computational methods with insufficient column relational information, AUTO-TUS directly model the complex column relational representations, which can better address the table union search task.
+
+<table><tr><td>Test / Train</td><td># Tables</td><td># Columns</td><td>Avg. # Rows</td><td>Size (GB)</td></tr><tr><td>SANTOS Small</td><td>550</td><td>6,322</td><td>6,921</td><td>0.45</td></tr><tr><td>TUS Small</td><td>1,530</td><td>14,810</td><td>4,466</td><td>1</td></tr><tr><td>TUS Large</td><td>5,043</td><td>54,923</td><td>1,915</td><td>1.5</td></tr><tr><td>SANTOS Small</td><td>550</td><td>6,162</td><td>7,732</td><td>0.5</td></tr><tr><td>TUS Small</td><td>1,530</td><td>13,926</td><td>4,032</td><td>0.9</td></tr><tr><td>TUS Large</td><td>5,043</td><td>55,462</td><td>1,729</td><td>1.3</td></tr></table>
+
+Table 1: The statistics of datasets for three benchmarks.
+
+# 4 Experimental Evaluation
+
+We conduct extensive experiments on real-world public datasets as well as a synthetic test set augmented with table noise to show the effectiveness of our AUTOTUS on table union search task, and give a detailed analysis.
+
+# 4.1 Experimental Setups
+
+Datasets. To evaluate our technique, three public datasets are used: SANTOS (Khatiwada and Fan, 2023), TUS small, and TUS large (Nargesian et al., 2018). These three datasets are subsets of Open Data released by authorities such as US Open Data $^{1}$ , UK Open Data $^{2}$ , and Canada Open Data $^{3}$  with manually labeled data. However, these data are usually officially filtered to remove a lot of table noise to maintain high quality. In order to verify the robustness of the model, we synthesized a test set with table noise, we added both column value and column name noises to the tabular data and obtained the synthesized SANTOS $_{N}$  and TUS large $_{N}$ . Note that although we add table noises, the human-annotated labels remain unchanged. To avoid information leakage, the manually labeled dataset is not used during the training. We randomly generated the same number of unlabeled training tables as the labeled tables from Open Data (Koutras et al., 2021; Khatiwada and Fan, 2023). We give the detailed statistics of datasets for three benchmarks in Table 1. We also give the dataset construction details in Appendix A.
+
+Experimental Settings and Metrics. We adopt BERT-Base, BERT-Large (Devlin et al., 2019) and RoBERTa-Large (Liu et al., 2019) as the encoder and set max-length as 512. For Column Relation Encoder Updater via Adaptive Clustering, we adopt  $k$ -means and set  $K = 50$  to get the initial centroids. We stop the clustering and classification loop when the current pseudo labels differs by less than  $10\%$  from the previous epoch. For Column
+
+Relation Encoder Updater via Pseudo Label Classification, we adopt the fully connected layer as  $\tau_{\phi}$  and set dropout rate as  $10\%$ , learning rate as  $1e - 5$  and warmup as 0.1. To allow fully connected layer to warm up, we fixed the parameters in  $f_{\theta}$  for the first two epochs. We add one noise of column data or column schemata to each column to obtain the training dataset. We first randomly choose whether to add noise in the column data or column schemata, and secondly, for all replacement, addition, or deletion of noise at the token level, we select  $20\%$  of all tokens in the cell of the column for modification.
+
+For evaluation metrics, following the previous work (Bogatu et al., 2020; Khatiwada and Fan, 2023), we adopt the Mean Average Precision at  $Y$  ( $MAP@Y$ ) and Recall at  $Y$  ( $R@Y$ ) to evaluate the effectiveness of the searched top- $Y$  table results. Note that  $MAP@Y$  is the average value of Precision at  $y$  ( $P@y$ ), where  $y = 1, 2, \dots, Y$ . Formally, given the query table  $Q$  and a set of data lake tables  $\mathcal{T}$ , we define  $\mathcal{T}_Q$  as the set of usable tables based on the ground truth and  $\mathcal{T}_Q'$  as the set of top- $Y$ 聢单table results using the searching methods. The  $P@Y$  and  $R@Y$  are calculated as:
+
+$$
+P @ Y = \frac {\mathcal {T} _ {Q} \cap \mathcal {T} _ {Q} ^ {\prime}}{\mathcal {T} _ {Q} ^ {\prime}}, R @ Y = \frac {\mathcal {T} _ {Q} \cap \mathcal {T} _ {Q} ^ {\prime}}{\mathcal {T} _ {Q}} \tag {10}
+$$
+
+Note that perfect  $R@Y$  is not possible when the ground truth contains less than  $Y$  searched table results. We define the Mean Average Precision  $MAP@Y$  as:
+
+$$
+M A P @ Y = \frac {1}{Y} \sum_ {y = 1} ^ {Y} P @ y \tag {11}
+$$
+
+For a fair comparison, we adopt the  $Y = 10$  on the SANTOS benchmarks and  $Y = 60$  on the TUS benchmarks to be consistent with Nargesian et al. (2018); Khatiwada and Fan (2023).
+
+Baseline Models. We compare AUTOTUS with baseline models in two categories. The first category is table union search model that adopts various column features to calculate the table unionability score. We adopt  $\mathbf{D}^3\mathbf{L}$  (Bogatu et al., 2020), SATO (Zhang et al., 2020), Starmie (Fan et al., 2022), and SANTOS (Khatiwada and Fan, 2023) as baselines. The second category is table pretrained model that leverages unlabeled tables for
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">SANTOS</td><td colspan="2">TUS small</td><td colspan="2">TUS large</td><td colspan="2">SANTOSN</td><td colspan="2">TUS largeN</td></tr><tr><td>MAP@10</td><td>R@10</td><td>MAP@60</td><td>R@60</td><td>MAP@60</td><td>R@60</td><td>MAP@10</td><td>R@10</td><td>MAP@60</td><td>R@60</td></tr><tr><td>BERT-Base (Devlin et al., 2019)</td><td>79.5</td><td>53.7</td><td>74.1</td><td>16.3</td><td>63.1</td><td>13.4</td><td>69.5</td><td>46.9</td><td>55.3</td><td>9.4</td></tr><tr><td>TaBERT (Yin et al., 2020)</td><td>83.6</td><td>56.2</td><td>80.3</td><td>18.7</td><td>76.4</td><td>15.8</td><td>74.8</td><td>50.7</td><td>67.3</td><td>11.4</td></tr><tr><td>TABBIE (Iida et al., 2021)</td><td>86.3</td><td>58.5</td><td>82.4</td><td>19.2</td><td>78.7</td><td>15.3</td><td>76.9</td><td>51.6</td><td>65.2</td><td>11.8</td></tr><tr><td>TUTA (Wang et al., 2021)</td><td>86.7</td><td>59.2</td><td>84.5</td><td>20.1</td><td>80.5</td><td>16.9</td><td>77.9</td><td>52.3</td><td>71.6</td><td>13.2</td></tr><tr><td>FORTAP (Cheng et al., 2022)</td><td>88.5</td><td>60.9</td><td>87.3</td><td>20.6</td><td>81.3</td><td>17.0</td><td>78.7</td><td>52.7</td><td>72.8</td><td>12.8</td></tr><tr><td>TableFormer (Yang et al., 2022)</td><td>89.4</td><td>61.4</td><td>86.9</td><td>20.2</td><td>83.2</td><td>17.3</td><td>79.2</td><td>53.5</td><td>72.7</td><td>12.9</td></tr><tr><td>D3L (Bogatu et al., 2020)</td><td>82.8</td><td>56.3</td><td>79.5</td><td>19.7</td><td>66.3</td><td>15.2</td><td>73.5</td><td>51.0</td><td>57.8</td><td>9.7</td></tr><tr><td>SATO (Zhang et al., 2020)</td><td>92.7</td><td>68.2</td><td>91.7</td><td>21.5</td><td>86.7</td><td>18.4</td><td>80.1</td><td>54.5</td><td>77.3</td><td>14.5</td></tr><tr><td>Starmie† (Fan et al., 2022)</td><td>93.9</td><td>69.0</td><td>94.8</td><td>23.3</td><td>89.5</td><td>18.9</td><td>81.0</td><td>55.1</td><td>78.4</td><td>14.3</td></tr><tr><td>SANTOS(Khatiwada and Fan, 2023)</td><td>94.3</td><td>69.2</td><td>85.7</td><td>21.0</td><td>-</td><td>-</td><td>81.4</td><td>55.3</td><td>-</td><td>-</td></tr><tr><td>AUTOTUSBERT-Base</td><td>98.5±0.4</td><td>72.4±0.2</td><td>97.7±0.5</td><td>25.2±0.3</td><td>94.6±0.7</td><td>21.6±0.2</td><td>87.6±0.5</td><td>59.7±0.3</td><td>84.9±0.6</td><td>17.3±0.3</td></tr><tr><td>w/o Table Noise Generator</td><td>96.6±0.5</td><td>71.2±0.4</td><td>96.1±0.6</td><td>24.5±0.3</td><td>92.5±0.7</td><td>20.8±0.4</td><td>83.0±0.7</td><td>56.4±0.4</td><td>80.7±0.8</td><td>14.9±0.3</td></tr><tr><td>w/o Encoder Update via Adaptive Clustering</td><td>94.5±0.6</td><td>71.9±0.4</td><td>94.3±0.7</td><td>23.3±0.5</td><td>91.0±0.5</td><td>20.4±0.3</td><td>83.2±0.6</td><td>56.3±0.4</td><td>80.5±0.7</td><td>14.7±0.3</td></tr><tr><td>w/o Encoder Update via Pseudo Label Classification</td><td>81.8±0.6</td><td>54.7±0.4</td><td>80.9±0.9</td><td>16.8±0.4</td><td>77.3±0.7</td><td>17.0±0.4</td><td>72.1±0.8</td><td>50.2±0.6</td><td>68.5±0.7</td><td>12.2±0.2</td></tr><tr><td>AUTOTUSBERT-Large</td><td>99.1±0.5</td><td>73.8±0.2</td><td>98.3±0.5</td><td>26.0±0.2</td><td>95.8±0.6</td><td>22.5±0.3</td><td>89.4±0.6</td><td>61.0±0.3</td><td>86.1±0.7</td><td>17.9±0.2</td></tr><tr><td>AUTOTUSRoBERTa-Large</td><td>99.4±0.4</td><td>73.9±0.2</td><td>98.9±0.3</td><td>26.7±0.2</td><td>96.2±0.5</td><td>22.3±0.2</td><td>90.0±0.5</td><td>60.5±0.3</td><td>86.9±0.6</td><td>18.1±0.3</td></tr></table>
+
+Table 2: MAP@Y and R@Y comparisons (%). Results of AUTOTUS are averaged over five runs.  $\dagger$ means we replace the base encoder from RoBERTa-Base to BERT-Base, and all models are based on BERT-Base for a fair comparison. Note that perfect R@Y is not possible when the ground truth contains less than Y searched tables.
+
+self-supervised pre-training and achieves promising results in various table related NLP tasks. We adopt TaBERT (Yin et al., 2020), TABBIE (Iida et al., 2021), TUTA (Wang et al., 2021), FORTAP (Cheng et al., 2022), and TableFormer (Yang et al., 2022) as baseline encoders to calculate the column unionability score and adopt our Table Scorer module to obtain the table unionability score. Although, there are more baseline models in this category, we just select representative and SOTA ones. The details of baselines are introduced in Appendix B.
+
+# 4.2 Results and Analysis
+
+Overall Performance. Table 2 shows the average results on five table union search benchmarks. The proposed AUTOUS consistently outperforms all baselines in MAP and R performance by leveraging column relational representation learning (with student's T test  $p < 0.05$ ). More specifically, for five benchmarks, compared to the SOTA baseline, AUTOUS on average achieves  $5.1\%$  higher MAP and  $3.1\%$  higher R across various benchmarks. Compared with BERT-Base, all table pre-trained and table union search baselines could gain MAP and R performance improvements from better learned tabular representations. The table union search baselines can achieve better results than the table pre-trained baselines by optimizing the column representations specific to the table union search task. AUTOUS exploit the column relational representations and are able to beat all the table union search baselines. In addition, we study the performance changes brought about by using larger base encoders BERT-Large (Devlin et al., 2019) and RoBERTa-Large (Liu et al., 2019).
+
+![](images/0499b61ea36c5f2efbfdfd3e7170cbf2cd14032c9fd6e44ce96056945e46f517.jpg)  
+(a)  $P@K$  on SANTOSN
+
+![](images/84830044b7884a84f423356c93293c564612bb0c77016960c91b2789c9f7b0c3.jpg)  
+(b)  $R@K$  on SANTOSN
+
+![](images/55eca785b1b56006f99c2d36356e329f33382d97f16379c93efa29292cd4d785.jpg)  
+(c)  $P@K$  on TUS Small
+
+![](images/f8c9a89fea0b3df1cdb4badc7095b352041011c75bfa6556a19a4a31e1cd09ff.jpg)  
+(d)  $R@K$  on TUS Small
+
+![](images/94ed419558c6f3d98114b4c62557686568eade91298fbd229872bc21699a48b1.jpg)  
+(e)  $P@K$  on TUS LargeN  
+Figure 3:  $\mathbb{P}@\mathcal{Y}$  and  $\mathbb{R}@\mathcal{Y}$  results on different datasets.
+
+![](images/bed15055a2cfe8be30caf286d0486fa7a211596e158b96136bc416126b186a4b.jpg)  
+(f)  $R@K$  on TUS LargeN
+
+From Figure 2, it can be concluded that leveraging a large-scale pre-trained language model with more parameters can lead to better MAP and R, BERT-Large and RoBERTa-Large obtain an average boost of  $1.1\%$  and  $1.4\%$ , respectively.
+
+Another interesting finding is that, although all the models have decreased performance on the datasets with table noise, the average performance improvement of AUTOTUS compared to SOTA is consistent, and even increased to  $6.6\%$  on MAP
+
+![](images/4ce794ff664341b5f89b03f1575dab91fb6d4c58cca9241404feed09b9016e15.jpg)  
+(a) AutoTUS w/o Column Pair Classification
+
+![](images/17cdd72e1216f58ac79e11d54b74163a72beaf9a7886f26e323beef0bbfb8487.jpg)  
+Figure 4: Visualizing column relational representations after t-SNE dimension reduction on TUS large $_N$ .
+
+![](images/6d1d1ed2a151aa00d423a9d5e1d7618d325b3008b2dbb787906ab8666f09cbdb.jpg)  
+(b) AutoTUS w/o Automatic Clustering (c)
+
+![](images/c255849864abb4d383a01f4e6235217e9c3b634527ae4f5d83b8ca9b8b570132.jpg)  
+AutoTUS w/o Real-world Noise Generation  
+(d) AutoTUS
+
+and  $3.8\%$  on R. We attribute the improvement of AUTCUS to the noisy, weak column relational representations from the large pre-trained language models are exploited and refined: we bootstrap the representations obtained from noisy tabular data via self-supervised training schema. More specifically, we give the comparison of  $\mathrm{P}@\mathrm{Y}$  and  $\mathrm{R}@\mathrm{Y}$  as  $Y$  changes in Figure 3. We find that AUTCUS consistently outperforms all baseline models on both metrics  $\mathrm{P}@\mathrm{Y}$  and  $\mathrm{R}@\mathrm{Y}$  as  $Y$  changes.
+
+Ablation Study. We conduct ablation studies to show the effectiveness of different modules of AUTOTUS. A number of variants of AUTOTUS are considered: AUTOTUS w/o Table Noise Generation removes added table noise in the training data; AUTOTUS w/o Column Relation Encoder Updater via Pseudo Label Classification is AUTOTUS without Column Relation Encoder Updater via Pseudo Label Classification and only uses the Contextualized Column Relation Encoder for Column Relation Encoder Updater via Adaptive Clustering; AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering replaces the suggested soft-assignment clustering techniques with  $k$ -means as a hard-assignment alternative.
+
+The general conclusion from ablation study results shown in Table 2 is that the all three modules contribute positively to improve the performance. More specifically, without table noise in training data, AUTOTUS w/o Table Noise Generation gives  $2.3\%$  less MAP and R averaged on all datasets, especially the noisy  $\mathrm{SANTOS}_N$  and TUS large, the drop reaches  $3.5\%$ . If we stop exploiting self-supervised signals for column relational features learning, AUTOTUS w/o Column Relation Encoder Updater via Pseudo Label Classification brings  $12.8\%$  less MAP and R averaged on all datasets. This huge performance drop fully demonstrates the importance of learning and refining column-pair representations. Column Relation Encoder Updater via Adaptive Clustering gives  $3.0\%$  MAP and R boost in average when comparing
+
+with the hard-assignment alternative (AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering).
+
+Visualizing Contextualized Column Relational Embeddings. To intuitively show how clustering-enhanced self-supervised training can exploit self-supervised signals to obtain better contextualized column relational representations, we visualize the dimensionally reduced column relational representation space  $\mathbb{R}^{2\cdot h_R}$  using t-SNE (Maaten and Hinton, 2008). We randomly choose 10 base tables from TUS large $_N$  dataset and sample all column relations that belong to the corresponding base table. We show the visualization results in Figure 4 with each column relation being colored according to their ground-truth.
+
+From Figure 4, we can see that AUTOTUS w/o Column Relation Encoder Updater via Pseudo Label Classification can assign meaningful semantics to column relations from different base tables, but these unrefined features cannot be tailored for the table union search task. When Column Relation Encoder Updater via Adaptive Clustering is not applied and only  $k$ -means are used, AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering performs a hard assignment of features, and the clustering results appear messy since there are no cluster centroids with high confidence. AUTOTUS w/o Table Noise Generation demonstrates that the clustered features is not tolerate to the noise samples, resulting in unclear cluster boundaries. AUTOTUS shows denser and well-separated clusters, which verifies its powerful column relational representation learning ability.
+
+Parameter Analysis: when  $\mathbf{K}$  is unknown. The Column Relation Encoder Updater via Adaptive Clustering provide the flexibility to explore column relational features without knowing any prior information about the number of clusters. This property is attractive when the number of target clusters is agnostic. We vary  $K$  from 10 to 200 and report the
+
+![](images/f8fe6fcada3e668877d0377794b95a349dfa73ad15b264734605fae713da175f.jpg)  
+AutoTUS (TUS Small) AutoTUS w/o Automatic Clustering (TUS Small)  
+AutoTUS (TUS Large) AutoTUS w/o Automatic Clustering (TUS Large)
+
+Figure 5: MAP@60 Score with different  $K$ .  
+
+<table><tr><td colspan="3">Table Noise</td><td colspan="2">SANTOSN</td><td colspan="2">TUS largeN</td></tr><tr><td>Column Name</td><td colspan="2">Column Data</td><td>MAP@10</td><td>R@10</td><td>MAP@60</td><td>R@60</td></tr><tr><td>X</td><td colspan="2">X</td><td>83.0</td><td>56.4</td><td>80.7</td><td>14.9</td></tr><tr><td>X</td><td colspan="2">✓</td><td>85.5</td><td>57.6</td><td>82.1</td><td>15.8</td></tr><tr><td>✓</td><td colspan="2">X</td><td>86.9</td><td>59.2</td><td>83.8</td><td>16.4</td></tr><tr><td>✓ or and</td><td colspan="2">✓</td><td>87.6</td><td>59.7</td><td>84.9</td><td>17.3</td></tr><tr><td></td><td></td><td>88.0</td><td>60.1</td><td>84.7</td><td>17.1</td><td></td></tr></table>
+
+Table 3: Table Noise Analysis.
+
+MAP@60 score in Figure 5, the best result is obtained when  $K = 25$  for TUS Small and  $K = 50$  for TUS Large, slightly larger than the number of base tables in the two datasets, indicating that AUTOTUS actually exploits the number of target clusters as useful prior knowledge. Thanks to the self-supervised training for column relational features exploitation and the flexibility brought by soft-assignment clustering, when we vary  $K$  from 10 to 200, AUTOTUS gets a more stable MAP@60 score than AUTOTUS w/o Column Relation Encoder Updater via Adaptive Clustering.
+
+Impact of Table Noise. We investigate the effect of table noise in column value and column name. As shown in Section 3.2, we randomly select a corresponding noise from two types of noises. From Table 3, we observe that adding table noise to both column value and column name help to improve the performance, whereas changing column name has a slightly larger effect (2.8% vs. 1.5%). This improvements may related to the fact that column name summarizes the entire column semantically, so adding noise to the column names can make the model more robust to table noise and thus obtain more performance gains. However, adding more noise did not consistently improve performance, suggesting that the model needs to find a trade-off between noisy data perturbation and increased robustness to noise.
+
+# 5 Related Work
+
+# 5.1 Table Union Search
+
+In dataset discovery, it is crucial to find related tables in data lakes. The table union search task has recently received considerable attentions (Ling et al., 2013; Lehmberg and Bizer, 2017; Khatiwada and Fan, 2023). The initial attempt is made by Nargesian et al. (2018). The  $\mathrm{D}^{3}\mathrm{L}$  (Bogatu et al., 2020) categories columns into groups by column features. Starmie (Fan et al., 2022) obtains the column features by utilizing a contextualized pretrained language models. SANTOS (Khatiwada and Fan, 2023) leverages a knowledge base to discover the unionable relations between two tables. More recently, a number of table pre-training models have been developed to obtain the contextualized table representation (Yin et al., 2020; Gong et al., 2020; Dong et al., 2022). However, these methods are not tailored to the table union search task.
+
+# 5.2 Self-supervised Learning in NLP
+
+Self-supervised learning (SSL) is a method for building models where the output labels are already included in the input data, eliminating the need for additional labeled data (Liu et al., 2021; Hu et al., 2021a,b; Liu et al., 2022d,c). SSL has been widely used in NLP domains such as sentence generation (West et al., 2019; Yan et al., 2021), document processing (You et al., 2021; Ginzburg et al., 2021), natural language inference (Li et al., 2022, 2023), and text reasoning (Klein and Nabi, 2020; Fu et al., 2020; Chen et al., 2022). BERT (Devlin et al., 2019) is one of the most eminent SSL methods which exploit self-supervisions from corpus with next sentence prediction and masked language modeling tasks. In our work, we adopt SSL method to exploit and refine self-supervised signals from tabular data.
+
+# 6 Conclusions
+
+In this paper, we propose a multi-stage self-supervised table union search framework, which represents column relation as column relational representation and learn this representation in a multi-stage manner that can better describe column relation for unionability prediction. In additional, a table noise generator is proposed to improve the robustness of our approach against the table noise. The conducted extensive experiments demonstrate the effectiveness of the proposed approach.
+
+# 7 Limitations
+
+We would like to claim our limitations from two perspectives: technical-wise and application-wise.
+
+Technical-wise: We currently only experiment with BERT-Base, BERT-Large, and RoBERTa-Large as the basic encoders. For larger language models, due to limited resources, we have not implemented them.
+
+Application-wise: The experimental data comes from the Open Data repository released by governments of various countries. Although many domains are covered, some domain-specific data, such as biomedical, have not been considered. Furthermore, our tabular data are all from English, open data research in other languages can be considered as a future research direction.
+
+# 8 Acknowledgement
+
+We thank the reviewers for their valuable comments. Lijie Wen is the corresponding author. Xuming Hu and Lijie Wen were partially supported by the National Key Research and Development Program of China (No. 2019YFB1704003), the National Nature Science Foundation of China (No. 62021002), Tsinghua BNrist and Beijing Key Laboratory of Industrial Bigdata System and Application. Philip S. Yu was partially supported by the NSF under grants III-1763325, III-1909323, III-2106758, SaTC-1930941.
+
+# References
+
+Rami Aly, Zhijiang Guo, Michael Sejr Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. Feverous: Fact extraction and verification over unstructured and structured information. In Proc. of NeurIPS: Datasets and Benchmarks Track.  
+Alex Bogatu, Alvaro AA Fernandes, Norman W Paton, and Nikolaos Konstantinou. 2020. Dataset discovery in data lakes. In Proc. of ICDE, pages 709-720. IEEE.  
+Dan Brickley, Matthew Burgess, and Natasha Noy. 2019. Google dataset search: Building a search engine for datasets in an open web ecosystem. In Proc. of WWW, pages 1365-1375.  
+Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020. Hybridqa: A dataset of multi-hop question answering over tabular and textual data. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1026-1036.
+
+Yubo Chen, Yunqi Zhang, and Yongfeng Huang. 2022. Learning reasoning patterns for relational triple extraction with mutual generation of text and graph. In *Findings of ACL*, pages 1638-1647.  
+Zhoujun Cheng, Haoyu Dong, Ran Jia, Pengfei Wu, Shi Han, Fan Cheng, and Dongmei Zhang. 2022. Fortap: Using formulas for numerical-reasoning-aware table pretraining. In Proc. of ACL, pages 1150-1166.  
+Nadiia Chepurko, Ryan Marcus, Emanuel Zgraggen, Raul Castro Fernandez, Tim Kraska, and David Karger. 2020. Arda: automatic relational data augmentation for machine learning. Proc. of VLDB, 13(9):1373-1387.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. of NAACL-HLT, pages 4171-4186.  
+Haoyu Dong, Zhoujun Cheng, Xinyi He, Mengyu Zhou, Anda Zhou, Fan Zhou, Ao Liu, Shi Han, and Dongmei Zhang. 2022. Table pretraining: A survey on model architectures, pretraining objectives, and downstream tasks. arXiv preprint arXiv:2201.09745.  
+Grace Fan, Jin Wang, Yuliang Li, Dan Zhang, and Renee Miller. 2022. Semantics-aware dataset discovery from data lakes with contextualized column-based representation learning. arXiv preprint arXiv:2210.01922.  
+Tsu-Jui Fu, Xin Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. 2020. Sscr: Iterative language-based image editing via self-supervised counterfactual reasoning. In Proc. of EMNLP, pages 4413-4422.  
+Sainyam Galhotra and Udayan Khurana. 2020. Semantic search over structured data. In Proc. of CIKM, pages 3381-3384.  
+Dvir Ginzburg, Itzik Malkiel, Oren Barkan, Avi Caci-ularu, and Noam Koenigstein. 2021. Self-supervised document similarity ranking via contextualized language models and hierarchical inference. In *Findings of ACL-IJCNLP*, pages 3088-3098.  
+Heng Gong, Yawei Sun, Xiaocheng Feng, Bing Qin, Wei Bi, Xiaojiang Liu, and Ting Liu. 2020. Tablegpt: Few-shot table-to-text generation with table structure reconstruction and content matching. In Proc. of COLING, pages 1978-1988.  
+Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178-206.  
+Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Mueller, Francesco Piccinno, and Julian Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. In Proc. of ACL, pages 4320-4333.
+
+Xuming Hu, Lijie Wen, Yusong Xu, Chenwei Zhang, and Philip S. Yu. 2020. Selfore: Self-supervised relational feature learning for open relation extraction. In Proc. of EMNLP, pages 3673-3682.  
+Xuming Hu, Chenwei Zhang, Fukun Ma, Chenyao Liu, Lijie Wen, and Philip S. Yu. 2021a. Semi-supervised relation extraction via incremental meta self-training. In Findings of EMNLP, pages 487-496.  
+Xuming Hu, Chenwei Zhang, Yawen Yang, Xiaohe Li, Li Lin, Lijie Wen, and Philip S. Yu. 2021b. Gradient imitation reinforcement learning for low resource relation extraction. In Proc. of EMNLP, pages 2737-2746.  
+Hiroshi Iida, Dung Thai, Varun Manjunatha, and Mohit Iyyer. 2021. Tabbie: Pretrained representations of tabular data. In Proc. of NAACL-HLT, pages 3446-3456.  
+Aamod Khatiwada and Grace Fan. 2023. Santos: Relationship-based semantic table union search. In Proc. of SIGMOD.  
+Tassilo Klein and Moin Nabi. 2020. Contrastive self-supervised learning for commonsense reasoning. In Proc. of ACL, pages 7517-7523.  
+Christos Koutras, George Siachamis, Andrea Ionescu, Kyriakos Psarakis, Jerry Brons, Marios Fragkoulis, Christoph Lofi, Angela Bonifati, and Asterios Katsifodimos. 2021. Valentine: Evaluating matching techniques for dataset discovery. In Proc. of ICDE, pages 468-479. IEEE.  
+Oliver Lehmberg and Christian Bizer. 2017. Stitching web tables for improving matching quality. Proc. of VLDB, 10(11):1502-1513.  
+Shu'ang Li, Xuming Hu, Li Lin, Aiwei Liu, Lijie Wen, and Philip S. Yu. 2023. A multi-level supervised contrastive learning framework for low-resource natural language inference. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:1771-1783.  
+Shu'ang Li, Xuming Hu, Li Lin, and Lijie Wen. 2022. Pair-level supervised contrastive learning for natural language inference. arXiv preprint arXiv:2201.10927.  
+Xiao Ling, Alon Y Halevy, Fei Wu, and Cong Yu. 2013. Synthesizing union tables from the web. In Proc. of IJCAI.  
+Aiwei Liu, Xuming Hu, Li Lin, and Lijie Wen. 2022a. Semantic enhanced text-to-sql parsing via iteratively learning schema linking graph. In Proc. of KDD, pages 1021-1030.  
+Aiwei Liu, Xuming Hu, Lijie Wen, and Philip S Yu. 2023. A comprehensive evaluation of chatgpt's zero-shot text-to-sql capability. arXiv preprint arXiv:2303.13547.
+
+Ruixue Liu, Shaozu Yuan, Aijun Dai, Lei Shen, Tiangang Zhu, Meng Chen, and Xiaodong He. 2022b. Few-shot table understanding: A benchmark dataset and pre-training baseline. In Proc. of COLING, pages 3741-3752.  
+Shuliang Liu, Xuming Hu, Chenwei Zhang, Shu'ang Li, Lijie Wen, and Philip S. Yu. 2022c. Hiure: Hierarchical exemplar contrastive learning for unsupervised relation extraction. In Proc. of NAACL-HLT, pages 5970-5980.  
+Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. 2021. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering.  
+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.  
+Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip Yu. 2022d. Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering.  
+Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(Nov):2579-2605.  
+Fatemeh Nargesian, Erkang Zhu, Ken Q Pu, and Renee J Miller. 2018. Table union search on open data. Proc. of VLDB, 11(7):813-825.  
+Aécio Santos, Aline Bessa, Christopher Musco, and Juliana Freire. 2022. A sketch-based index for correlated dataset search. In Proc. of ICDE, pages 2928-2941. IEEE.  
+Livio Baldini Soares, Nicholas Fitzgerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In Proc. of ACL, pages 2895-2905.  
+Zhiruo Wang, Haoyu Dong, Ran Jia, Jia Li, Zhiyi Fu, Shi Han, and Dongmei Zhang. 2021. Tuta: tree-based transformers for generally structured table pretraining. In Proc. of SIGKDD, pages 1780-1790.  
+Peter West, Ari Holtzman, Jan Buys, and Yejin Choi. 2019. Bottlesum: Unsupervised and self-supervised sentence summarization using the information bottleneck principle. In Proc. of EMNLP-IJCNLP, pages 3752-3761.  
+Yihai Xi, Ning Wang, Shuang Hao, Yiyi Zhang, and Xinyu Chen. 2023. Popularity sensitive and domain-aware summarization for web tables. Information Sciences, 621:729-748.  
+Junyuan Xie, Ross Girshick, and Ali Farhadi. 2016. Unsupervised deep embedding for clustering analysis. In Proc. of ICML, pages 478-487.
+
+Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. 2021. Consent: A contrastive framework for self-supervised sentence representation transfer. In Proc. of ACL-IJCNLP, pages 5065-5075.  
+Jingfeng Yang, Aditya Gupta, Shyam Upadhyay, Luheng He, Rahul Goel, and Shachi Paul. 2022. Tableformer: Robust transformer modeling for table-text encoding. In Proc. of ACL, pages 528-537.  
+Pengcheng Yin, Graham Neubig, Wen-tau Yih, and Sebastian Riedel. 2020. Tabert: Pretraining for joint understanding of textual and tabular data. In Proc. of ACL, pages 8413-8426.  
+Chenyu You, Nuo Chen, and Yuexian Zou. 2021. Self-supervised contrastive cross-modality representation learning for spoken question answering. In *Findings of EMNLP*, pages 28-39.  
+Dan Zhang, Madelon Hulsebos, Yoshihiko Suhara, Cagatay Demiralp, Jinfeng Li, and Wang-Chiew Tan. 2020. Sato: contextual semantic type detection in tables. Proceedings of the VLDB Endowment, 13(12):1835-1848.  
+Shuo Zhang and Krisztian Balog. 2017. Entitables: Smart assistance for entity-focused tables. In Proc. of SIGIR, pages 255-264.  
+Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. 2021. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. In Proc. of ACL-IJCNLP, pages 3277-3287.
+
+# A Dataset Construction Details
+
+For test data, the SANTOS Small benchmark (Khatiwada and Fan, 2023) consists of 550 data lake tables, which are generated from 296 open datasets from Canada, the United Kingdom, the United States, and Australia. The SANTOS Small also has 50 query tables. There are two benchmarks accessible from Nargesian et al. (2018): TUS Small and TUS Large. The TUS Small consists of 1,530 data lake tables that are generated from 10 base table of Canada open data. The TUS Small also has 150 query tables. The TUS Large consists of 5,043 data lake tables that are generated from 32 base table of Canada open data. The TUS Large also has 100 query tables. The SANTOS $^{4}$  and TUS $^{5}$  benchmarks and their unionable table ground truth are publicly available. For training data, we randomly generated the same number of unlabeled training tables as the labeled tables from base tables of Open Data. Note that the public code for generating tables is publicly available $^{6}$  (Koutras et al., 2021; Khatiwada and Fan, 2023). Since table union search task is an unsupervised task, no dev data is obtained.
+
+# B The Introduction of Baseline Models
+
+We compare AUTOTUS with two categories of baseline models. The first category of models leverage unlabeled tables for self-supervised pretraining and achieve promising results in the table understanding tasks as baseline encoders to calculate the column unionability score:
+
+(1) TaBERT (Yin et al., 2020) is a pretrained language model that simultaneously learns representations for (semi-)structured tables and natural language phrases. 26 million tables and their English contexts make up the vast corpus on which TaBERT was trained.  
+(2) TABBIE (Iida et al., 2021) develops a straightforward pretraining target (corrupt cell identification) that only learns from tabular data and achieves the state-of-the-art on the table-based tasks. TABBIE offers embeddings of all table substructures (cells, rows, and columns), unlike competing techniques, and it also takes far less computing power to train.
+
+(3) TUTA (Wang et al., 2021) is a unified pretraining architecture for comprehending typically arranged tables. TUTA improves transformers with three structure-aware techniques after realizing that understanding a table necessitates spatial, hierarchical, and semantic information.  
+(4) FORTAP (Cheng et al., 2022) explores to leverage the spreadsheet formulas for table pretraining. FORTAP adopts two self-supervised pretraining objectives, which are derived from formulas, numerical reference prediction and numerical calculation prediction.  
+(5) TableFormer (Yang et al., 2022) is a structurally conscious table-text encoding architecture in which learnable attention biases are used to fully include tabular structural biases.
+
+Of course, there are more baseline models in this category, we just select representative and SOTA models. The second category of models adopt various column representations to calculate the column unionability score:
+
+(6)  $\mathbf{D}^3\mathbf{L}$  (Bogatu et al., 2020) creates hash-based indexes using the features of the items in a dataset and maps those features to a uniform distance space.  
+(7) SATO (Zhang et al., 2020) is a hybrid machine learning model that uses both the context of the table and the values of the columns to automatically identify the semantic categories of columns in tables.  
+(8) Starmie (Fan et al., 2022) obtains the semantic information included inside tables by utilizing a contrastive multi-column pre-training technique.  
+(9) SANTOS (Khatiwada and Fan, 2023) suggests a definition of unionability that takes connections between columns and their semantics into principled consideration.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? Section 7  
+A2. Did you discuss any potential risks of your work? Section 7  
+A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 3 and Section 4
+
+B1. Did you cite the creators of artifacts you used? Section 3 and Section 4  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3 and Section 4  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3 and Section 4  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Section 3, Section 4, and Appendix A  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4, and Appendix A
+
+C Did you run computational experiments?
+
+Section 4
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4 and Appendix A  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.

2023/Automatic Table Union Search with Tabular Representation Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa55852db4533ef4b283836b2866d41db2480f4410a2797eefc95348d9a87ef1
+size 491069

2023/AxomiyaBERTa_ A Phonologically-aware Transformer Model for Assamese/86b88357-8c3a-4270-bb19-cd2104e1901b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a96509f0669efc0cd1badab43aded9d07837569068322749cbc2537b9cd864c8
+size 844734

2023/AxomiyaBERTa_ A Phonologically-aware Transformer Model for Assamese/full.md

@@ -0,0 +1,467 @@
+# AxomiyaBERTa: A Phonologically-aware Transformer Model for Assamese
+
+Abhijnan Nath, Sheikh Mannan, and Nikhil Krishnaswamy
+
+Situated Grounding and Natural Language (SIGNAL) Lab
+
+Department of Computer Science, Colorado State University
+
+Fort Collins, CO, USA
+
+{abhijnan.nath, sheikh.mannan, nkrishna}@colostate.edu
+
+# Abstract
+
+Despite their successes in NLP, Transformer-based language models still require extensive computing resources and suffer in low-resource or low-compute settings. In this paper, we present AxomiyaBERTa, a novel BERT model for Assamese, a morphologically-rich low-resource language (LRL) of Eastern India. AxomiyaBERTa is trained only on the masked language modeling (MLM) task, without the typical additional next sentence prediction (NSP) objective, and our results show that in resource-scarce settings for very low-resource languages like Assamese, MLM alone can be successfully leveraged for a range of tasks. AxomiyaBERTa achieves SOTA on token-level tasks like Named Entity Recognition and also performs well on "longer-context" tasks like Cloze-style QA and Wiki Title Prediction, with the assistance of a novel embedding disperser and phonological signals respectively. Moreover, we show that AxomiyaBERTa can leverage phonological signals for even more challenging tasks, such as a novel cross-document coreference task on a translated version of the ECB+ corpus, where we present a new SOTA result for an LRL. Our source code and evaluation scripts may be found at https://github.com/CSU-signal/axomiyaberta.
+
+# 1 Introduction
+
+Transformer-based neural architectures such as BERT (Devlin et al., 2019) have revolutionized natural language processing (NLP). The ability to generate contextualized embeddings that both preserve polysemous word sense and similarity across dimensions through self-attention has contributed to significant improvements in various NLP tasks (Ethayarajh, 2019). Despite their successes, Transformers come at a high computational cost (Zhao et al., 2022) and still suffer from long-standing issues pertaining to data-hunger and
+
+availability of training resources. One effect of the dependency on big data is the continued proliferation of sophisticated NLP for well-resourced languages while low-resourced languages (LRLs) continue to be underrepresented, and the disparities continue to grow (Joshi et al., 2020).
+
+This is particularly true for languages of India and South Asia where English is widely spoken among the educated and urban population. Therefore, those in India most likely to use and develop NLP may freely do so in English, but sole speakers of local Indian languages may remain effectively isolated from human language technology in their native tongues. While strides have been made in NLP for widely-spoken Indian languages (e.g., Hindi, Bengali, Marathi, Tamil, etc.), India is home to about a thousand languages, over 100 of which are considered "major" but are not widely represented in NLP research. This lack of representation also precludes insights from those languages from contributing to the field (Bender, 2019).
+
+In this paper, we present AxomiyaBERTa, a novel Transformer language model for the Assamese language.2 AxomiyaBERTa has been trained in a low-resource and limited-compute setting, using only the masked language modeling (MLM) objective. Beyond a model for a new language, our novel contributions are as follows:
+
+- Use of a novel combined loss technique to disperse AxomiyaBERTa's embeddings;  
+- Addition of phonological articulatory features as an alternate performance improve
+
+ment in the face of omitting the NSP training objective for longer-context tasks;
+
+- Evaluation on event coreference resolution, which is novel for Assamese.
+
+AxomiyaBERTa achieves competitive or state of the art results on multiple tasks, and demonstrates the utility of our approach for building new language models in resource-constrained settings.
+
+# 2 Related Work
+
+Multilingual large language models (MLLMs) trained over large Internet-sourced data, such as MBERT and XLM (Conneau et al., 2020), provide resources for approximately 100 languages, many of which are otherwise under-resourced in NLP. However, multiple publications (Virtanen et al., 2019; Scheible et al., 2020; Tanvir et al., 2021) have demonstrated that multilingual language models tend to underperform monolingual language models on common tasks; the "multilingual" quality of MLLMs may not be enough to assure performance on LRL tasks, due to language-specific phenomena not captured in the MLLM.
+
+Since languages that share recent ancestry or a Sprachbund tend to share features, there has also been development of models and resources for languages from distinct regions of the world. South Asia is one such "language area," where even unrelated languages may share features (e.g., 4-way voice/ascpiration distinctions, SOV word order, retroflex consonants, heavy use of light verbs). As such, researchers have developed region-specific models for South Asian languages such as IndicBERT (Kakwani et al., 2020) (11 languages, 8.8 billion tokens) and MuRIL (Khanuja et al., 2021) (17 languages, 16 billion tokens).
+
+Subword tokenization techniques like byte-pair encoding (BPE) (Sennrich et al., 2016) yield comparatively better performance on LRLs by not biasing the vocabulary toward the most common words in a specific language, but BPE tokens also further obscure morphological information not immediately apparent in the surface form of the word. Nzeyimana and Niyongabo Rubungo (2022) tackle this problem for Kinyarwanda using a morphological analyzer to help generate subwords that better capture individual morphemes. However, despite similar morphological richness of many Indian languages, and likely due to similar reasons as outlined above, the dearth of NLP technology for most Indian languages extends to a lack of morphological parsers. We hypothesize that adding
+
+phonological features can also capture correlations between overlapping morphemes.
+
+Previous NLP work in Assamese includes studies in corpus building (Sarma et al., 2012; Laskar et al., 2020; Pathak et al., 2022), POS tagging (Kumar and Bora, 2018), WordNet (Bharali et al., 2014; Sarmah et al., 2019) structured representations (Sarma and Chakraborty, 2012), image captioning (Nath et al., 2022c), and cognate detection (Nath et al., 2022a). There does not exist, to our knowledge, significant work on Assamese distributional semantics, or any monolingual, Transformer-based language model for the Assamese language evaluated on multiple tasks.
+
+Our work complements these previous lines of research with a novel language model for Assamese, which further develops an initial model first used in Nath et al. (2022a). We account for the lack of an Assamese morphological analyzer with additional phonological features and task formulations that allow for strategic optimization of the embedding space before the classification layer.
+
+# 2.1 Assamese
+
+Assamese is an Eastern Indo-Aryan language with a speaker base centered in the Indian state of Assam. It bears similarities to Bengali and is spoken by 15 million L1 speakers (up to 23 million total speakers). Its literature dates back to the 13th c. CE. It has been written in its modern form since 1813, is one of 22 official languages of the Republic of India, and serves as a lingua franca of the Northeast Indian region (Jain and Cardona, 2004).
+
+Despite this, Assamese data in NLP resources tends to be orders of magnitude smaller than data in other languages, even in South Asian region-specific resources (see Table 1).
+
+<table><tr><td></td><td>as</td><td>bn</td><td>hi</td><td>en</td></tr><tr><td>CC-100</td><td>5</td><td>525</td><td>1,715</td><td>55,608</td></tr><tr><td>IndicCorp</td><td>32.6</td><td>836</td><td>1,860</td><td>1,220</td></tr></table>
+
+Table 1: CC-100 (Conneau et al., 2020) and IndicCorp (Kakwani et al., 2020) data sizes (in millions of tokens) for Assamese, Bengali, Hindi, and English.
+
+Assamese bears a similar level of morphological richness to other Indo-Aryan and South Asian languages, with 8 grammatical cases and a complex verbal morphology. Despite these points of comparison, Assamese has some unique phonological features among Indo-Aryan languages, such as the use of alveolar stops  $/t^{(h)}/, /d^{(f)}/,$  velar
+
+fricative  $/x/$ , and approximant  $/\mathrm{I}/$ . This atypical sound pattern motivates the use of phonological signals in our model. Moreover, both the pretraining and task-specific corpora we use contain a large proportion of loanwords (e.g., from English) or words cognate with words in higher-resourced languages (e.g., Bengali). These words rendered with Assamese's unique sound pattern result in distinct, information-rich phoneme sequences.
+
+# 3 Methodology
+
+# 3.1 Pretraining
+
+We trained on four publicly-available Assamese datasets: Assamese Wikidumps<sup>3</sup>, OS-CAR (Suárez et al., 2019)<sup>4</sup>, PMIndia (Haddow and Kirefu, 2020)<sup>5</sup>, the Common Crawl (CC-100) Assamese corpus (Conneau et al., 2020)<sup>6</sup>, as well as a version of the ECB+ Corpus (Cybul-ska and Vossen, 2014) translated to Assamese using Microsoft Azure Translator. In total, after preprocessing, the training data amounts to approximately 26 million space-separated Assamese tokens.<sup>7</sup>
+
+AxomiyaBERTa (66M parameters) was trained as a "light" ALBERT (specifically albert-base-v2) (Lan et al., 2019) model with only the MLM objective (Devlin et al., 2019), and no next sentence prediction (NSP), for 40 epochs (485,520 steps) with a vocabulary size of 32,000 and a SentencePiece BPE tokenizer (Kudo and Richardson, 2018). Tokenization methods like BPE can obfuscate certain morphological information. However, without a publicly-available morphological analyzer for Assamese, our motivation was to examine if phonological correlations might pick up similar information across different tasks while keeping model architecture and tokenizer consistent. We trained on 1 NVIDIA A100 80 GB device with a batch size of 32 and a sequence length of 128 for approximately 72 hours. Table 8 in Appendix A shows all specific pretraining configuration settings.
+
+# 3.1.1 Special Token Vocabulary
+
+The AxomiyaBERTa vocabulary includes two special trigger tokens:  $\langle \mathfrak{m} \rangle$  and  $\langle \mathfrak{m} \rangle$ . These act as separators  $a \, la$  the BERT [SEP] token, meaning that contextualized representations of these tokens were trained into the AxomiyaBERTa embedding space. Prior to pretraining, the translated ECB+ Corpus was annotated with these tokens surrounding event mentions. Since AxomiyaBERTa was not trained using the next sentence prediction objective (see Sec. 3.2.2), its embedding space needs those special triggers as separators between segments instead of the [SEP] tokens that segregate the token type IDs.
+
+# 3.2 Fine-tuning
+
+AxomiyaBERTa pretraining created a task-agnostic model optimized for the grammar and structure of Assamese. This model was then fine-tuned to achieve good performance on a number of different tasks. Beyond the task-specific fine-tuning, we made use of two auxiliary techniques: an embedding disperser, that optimized the AxomiyaBERTa embedding space away from severe anisotropy, and phonological or articulatory attention that acted as a single-head attention layer attending to both token-level and candidate-option level phonological signals. We first discuss these two techniques, followed by the specific task formulations we evaluated on. Note that the embedding disperser was used at the fine-tuning stage for Cloze-QA only due to severe anisotropy of the embedding space (Fig. 1 and Fig. 4, Appendix B).
+
+# 3.2.1 Embedding Disperser
+
+Without a meaningful objective to force embedding vectors apart during training, they trend toward an arbitrary center in  $\mathbb{R}^d$  space. This phenomenon has also been observed by Gao et al. (2018), Ethayarajh (2019), and Demeter et al. (2020), among others. In Nath et al. (2022a), evidence was presented that the effect is more pronounced in smaller models. An effect of this can be illustrated by embeddings from an example task, Cloze-style question answering (Cloze-QA):
+
+Let a "set" of embeddings consist of representations for a question (or context)  $Q$  and associated candidate answers  $\{A,B,C,D\}$ . "Within-set" cosine similarities represent the cosine similarities between  $(Q + i,Q + j)$  for each candidate answer  $i\in \{A,B,C,D\}$  and each other candid
+
+![](images/45b16f4335f5e4fc57db8606164b89f59186b2287adf3833ee8cd214f443ab4f.jpg)  
+Figure 1: Kernel Density Estimation plots for within-set features of each component of the phonologically-aware embedding disperser (Fig. 2). option_cos is the output of the auxiliary discriminator while linear_sigmoid represents the linear layer and euc_dist represents  $\mathrm{L}^2$  norm between the raw [CLS] token embeddings. See Fig. 4 in Appendix B for equivalent "beyond-set" plots.
+
+date  $j \in \{A, B, C, D\}$  where  $i \neq j$ . "Beyond-set" cosine similarities represent similarities between all pairs in a candidate-plus-answers set compared to other such embedding sets from different questions. Fig. 1 shows KDE plots for various similarity metrics taken "within-set" for a random sample of 100 sets from the Cloze-QA dev set (see Sec. 3.2.3 for more details on the data). The blue spike at 1 for cls_cosine_sim shows how similar all [CLS] token embeddings are to each other, given AxomiyaBERTa's extremely anisotropic embedding space after pretraining. This makes it difficult to optimize a classification boundary during fine-tuning using standard techniques.
+
+Therefore, to disperse the embedding space for greater discriminatory power, we used a combination of Binary Cross Entropy loss and Cosine Embedding loss to train the model. The architecture is shown in Fig. 2. The key components are: i) a cosine embedding layer that takes in arg1 (context) and arg2 (candidate) representations along with a [CLS] representation and outputs a 128D embedding into the cosine embedding loss function, and ii) an auxiliary discriminator that considers only arg2 and [CLS] representations.
+
+Mathematically,
+
+$$
+L _ {B C E} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \left(Y _ {i} \cdot \log \hat {Y} _ {i} + (1 - Y _ {i}) \cdot \log \left(1 - \hat {Y} _ {i}\right)\right)
+$$
+
+$$
+L _ {C O S} (x, y) = \left\{ \begin{array}{l l} 1 - \cos \left(x _ {1}, x _ {2}\right), & \text {i f} y = 1 \\ \max  \left(0, \cos \left(x _ {1}, x _ {2}\right) - \mathrm {m}\right), & \text {i f} y = - 1 \end{array} \right.
+$$
+
+$$
+L _ {C O M B} = \alpha L _ {B C E} + L _ {C O S} (x, y)
+$$
+
+where  $m$  represents the margin for the cosine loss and  $\alpha$  is 0.01.  $x_{1}$  corresponds to  $\arg 1$  and  $x_{2}$  corresponds to  $\arg 2$ .  $y = 1$  if  $x_{2}$  is the correct answer and  $y = -1$  if not. At inference, we computed Euclidean distance between the embedding outputs of the auxiliary discriminator and the cosine embedding layer with a threshold  $T$  of 4.45, found through hyperparameter search.
+
+option_cosine_sim in Fig. 1 shows the outputs of the embedding disperser's cosine embedding layer while option.cos shows the outputs of the auxiliary discriminator. In both cases we see distinct distributions that separate correct and incorrect answers. Works such as Cai et al. (2021) present evidence of such cases of global token anisotropy in other Transformer models and suggest that creating such local isotropic spaces leads to better results in downstream tasks.
+
+# 3.2.2 Phonological/Articulatory Attention
+
+While the NSP objective is effective at training LLMs to encode long-range semantic coherence (Shi and Demberg, 2019), it comes at a significant additional computational cost. Moreover, for very low-resource languages like Assamese, a lack of available long document or paragraph data means there may not exist a sufficient volume of coherent consecutive sentences in the training data.
+
+We hypothesize that when fine-tuning a smaller model like AxomiyaBERTa in a resource-constrained setting, adding phonological signals to the latent representations of text samples allows us to achieve a balanced trade-off between possible information loss due to reduced supervision (no NSP objective) and improved task-specific performance, at a lower compute cost.
+
+Previous works (e.g., Mortensen et al. (2016); Rijhwani et al. (2019); Nath et al. (2022b)) have shown that phonological features are useful for both token-level "short-context" tasks like NER or loanword detection as well as "longer-context" tasks like entity linking. We fine-tune for longer-context tasks by encoding candidate answers as phonological features and the pooled embedding of the context, and computing the relative difference in mutual information between each candidate answer and the context. High variance in cosine similarities within pairs in a context-candidate set is due to the phonological signals.
+
+![](images/286385357a94f434079fede8d3f9164134b46d801948542a90d5bf985ac5adc9.jpg)  
+Figure 2: Embedding disperser architecture with Cosine Embedding and Binary Cross Entropy (BCE) Loss. ClozeQA data used as example.
+
+Table 2 shows that the mean, standard deviation, and variances of [CLS] token cosine similarities for pretrained AxomiyaBERTa are much smaller than those extracted from XLM, but fine-tuning with phonological signals brings AxomiyaBERTa's values much closer XLM's.
+
+<table><tr><td></td><td>AxB</td><td>XLM</td><td>AxB + Phon</td></tr><tr><td>Mean</td><td>.998</td><td>.82</td><td>.67</td></tr><tr><td>Variance</td><td>5e-6</td><td>.08</td><td>.06</td></tr><tr><td>Stdev</td><td>.002</td><td>.28</td><td>.25</td></tr><tr><td>Min</td><td>.993</td><td>.13</td><td>.17</td></tr></table>
+
+Table 2: Statistics of AxomiyaBERTa (AxB) and XLM's [CLS] token cosine similarities compared to those of the pooled output of AxomiyaBERTa with phonological signals (AxB + Phon) over 100 random samples of within-set pairs of Cloze-QA dev set.
+
+To extract phonological features, we used the Assamese grapheme-to-phoneme mapping from Nath et al. (2022a), written for the Epitran library (Mortensen et al., 2018)<sup>8</sup> to convert all text into the International Phonetic Alphabet (IPA). We then used the PanPhon library (Mortensen et al., 2016) to convert the IPA transcriptions into 24 subsegmental features such as place and manner of articulation, voicing, etc.
+
+These feature vectors are padded to the maximum length (across train, test, and dev sets), and then concatenated to either the pooled context embedding (for long-context tasks) or the named-entity token embedding (for NER tasks).
+
+# 3.2.3 Cloze-style multiple-choice QA
+
+We fine-tuned AxomiyaBERTa on the Cloze-style Wiki question answering task from the IndicGLUE dataset (Kakwani et al., 2020). We surrounded both the masked text segment as well as the four candidate answers with the special tokens  $(<\mathfrak{m}>)$  and  $\langle \mathfrak{m} \rangle$  and then fed them into the pretrained AxomiyaBERTa model to get pairwise scores with BCE loss. Positive samples were labeled as 1 and negatives as 0. The encoded representation for each sample was a concatenation of the pooled ([CLS]) token output, the averaged embedding for the masked text segment ( $\arg 1$ ), that of the candidate answer ( $\arg 2$ ), and the element-wise multiplication of  $\arg 1$  and  $\arg 2$ . This was input into a pairwise scorer  $a la Caci-ularu$  et al. (2021). We fine-tuned our model (with and without phonological attention) with the pairwise scorer head for 5 iterations with a batch size of 80, a scorer head learning rate of  $1e-4$  and a model learning rate of  $2e-5$ .
+
+# 3.2.4 Named Entity Recognition (NER)
+
+For NER, we fine-tuned and evaluated AxomiyaBERTa on two datasets: WikiNER (Pan et al., 2017) and AsNER (Pathak et al., 2022). For both datasets, we fed in the tokenized sentence while masking out all sub-word tokens except the first of each word. We used a token-classification head fine-tuned using a multi-class cross-entropy loss for the label set of the respective datasets. For our model without phonological signals, we fine-tuned for 10 epochs with a learning rate of 2e-5 with a linear LR scheduler and a batch size of 20. For our phonological attention-based model, we fine-tuned for 20 epochs with a batch size of 40 while keeping all other hyperparameters the same.
+
+# 3.2.5 Wikipedia Section Title Prediction
+
+Like Cloze-QA, this task comes from IndicGLUE (Kakwani et al., 2020). Fine-tuning for this task was quite similar to that of Cloze-QA, except we did not surround the candidates or the contexts with the trigger tokens. We fed in the Wikipedia section text and candidate title and optimized the multi-class cross entropy loss with a multiple choice head. We fine-tuned for 20 epochs with a batch size of 40. For the phonologically-aware model, we concatenated the articulatory signals to the pooled embedding output for each sample and fine-tuned our model for 200 iterations with a batch size of 40. We used a smaller model learning rate of 1e-6 and a classifier head learning rate of 9.5e-4 for both these models.
+
+# 3.2.6 Pairwise Scorer for Assamese CDCR
+
+Coreference resolution in a cross-document setting (DCRC) involves identifying and clustering together mentions of the same entity across a set of documents (Lu and Ng, 2018). Following CDCR approaches in Cattan et al. (2021) and Caciularu et al. (2021), we trained a pairwise scorer with BCE loss over all antecedent spans for each sentence containing an event (across all documents) while ignoring identical pairs. We generated concatenated token representations from Transformer-based LMs by joining the two paired sentences after surrounding the event mentions with the special trigger tokens. These representations were input to the pairwise scorer (PS) to calculate affinity scores between all those pairs. Mathematically,
+
+$$
+\operatorname {S c o r e s} (i, j) = P S ([ C L S ], f (x), f (y), f (x) * f (y)),
+$$
+
+where  $[CLS]$  represents the pooled output of the entire sentence pair,  $f(x)$  and  $f(y)$  are the representations of the two events (in context) and  $*$  represents element-wise multiplication.
+
+We trained the Pairwise Scorer for 10 epochs for all baseline models as well as AxomiyaBERTa. At inference, we used a connected-components clustering technique with a tuned threshold to find coreferent links. For baselines and ablation tasks, we calculated coreference scores using a lemmabased heuristic, and fine-tuned four other popular MLLMs using the same hyperparameters. More details and analysis are in Appendix D.
+
+# 4 Evaluation
+
+Table 3 shows the number of samples in the train, dev, and test splits, and the padding length, for all
+
+tasks we evaluated on. For Cloze-QA and WikiTitles, we evaluated on IndicGLUE. For NER, we evaluated on AsNER and WikiNER. For our novel coreference task, we evaluated on the translated  $\mathrm{ECB + }$  corpus, where the ratio of coreferent to noncoreferent pairs in the test set is approximately 1:35. We conducted exhaustive ablations between native and the phonologically-aware models for each task, and compared to previously-published baselines where available. For Cloze-QA, we created a train/test split of approximately 4.5:1. We fine-tuned off-the-shelf IndicBERT and MBERT on AsNER for 10 epochs on 1 NVIDIA RTX A6000 48 GB device with a batch size of 20.
+
+<table><tr><td>Features</td><td>Train</td><td>Dev</td><td>Test</td><td>Pad-Len</td></tr><tr><td>Cloze-QA</td><td>8,000</td><td>2,000</td><td>1,768</td><td>360</td></tr><tr><td>Wiki-Titles</td><td>5,000</td><td>625</td><td>626</td><td>1,848</td></tr><tr><td>AsNER</td><td>21,458</td><td>767</td><td>1,798</td><td>744</td></tr><tr><td>WikiNER</td><td>1,022</td><td>157</td><td>160</td><td>480</td></tr><tr><td>T-ECB+</td><td>3,808</td><td>1,245</td><td>1,780</td><td>552</td></tr></table>
+
+Table 3: Table showing distribution of train/dev/test splits for all tasks. T-ECB+ signifies number of event mentions in the translated ECB+ corpus, keeping special trigger tokens in place. "Pad-Len" represents the maximum padded length of the articulatory feature embeddings generated from PanPhon for all three splits.
+
+# 5 Results and Discussion
+
+Table 4 shows Test F1 Scores/Accuracy for AxomiyaBERTa for the various short-context (classification) and long-context (multiple-choice) tasks. We compared baselines from previous works and newly fine-tuned baselines for certain tasks. We used the same pretrained model for all experiments with task fine-tuning heads consistent with previous benchmarks (Kakwani et al., 2020). One exception is the Cloze-QA task where we dealt with task-specific severe anisotropy with embedding dispersal.
+
+# 5.1 Short-context: AsNER and WikiNER
+
+AxomiyaBERTa achieved SOTA performance on the AsNER task and outperformed most other Transformer-based LMs on WikiNER.
+
+Phonologically-aware AxomiyaBERTa Our experiments suggest that phonological signals are informative additional features for short-context tasks like NER for low-resourced, smaller models like AxomiyaBERTa. Table 4 shows that phonologically-aware AxomiyaBERTa outperformed non-phonological (hereafter "native")
+
+<table><tr><td>Models</td><td>Cloze-QA</td><td>Wiki-Titles</td><td>AsNER (F1)</td><td>WikiNER (F1)</td></tr><tr><td>XLM-R</td><td>27.11</td><td>56.96</td><td>69.42</td><td>66.67</td></tr><tr><td>MBERT</td><td>29.42</td><td>73.42</td><td>68.02*</td><td>92.31</td></tr><tr><td>IndicBERT-BASE</td><td>40.49</td><td>65.82</td><td>68.37*</td><td>41.67</td></tr><tr><td>MuRIL</td><td>-</td><td>-</td><td>80.69</td><td>-</td></tr><tr><td>AxomiyaBERTa</td><td>46.66</td><td>26.19</td><td>81.50</td><td>72.78</td></tr><tr><td>AxomiyaBERTa + Phon</td><td>47.40</td><td>59.26</td><td>86.90</td><td>81.71</td></tr></table>
+
+Table 4: Test F1 Scores/Accuracy for AxomiyaBERTa on all evaluation tasks, compared to previous baselines and our fine-tuned baselines. "AxomiyaBERTa + Phon" shows results for phonologically-aware AxomiyaBERTa. AsNER scores with a * represent versions we fine-tuned for this task. For Cloze-QA, Wiki-Titles and WikiNER, other model performances are from Kakwani et al. (2020). Bold indicates best performance.
+
+![](images/a4249c28cefc5e7d5f7947d3eb42b82111f84b983ba0ec3b084453797e1a4607.jpg)
+
+![](images/c236f6c1180aa188e13b9b9b84d0710b189bc7749b41f252aa5851eaabd0f7cf.jpg)
+
+![](images/7fa811159dfaf661969e0a06d32f06c8525761f5d9d78ebb6cb705559ee9c7e4.jpg)  
+Figure 3: Top: Confusion matrices showing AxomiyaBERTa performance on AsNER without [L] and with [R] phonological awareness. Bottom: IndicBERT [L] and MBERT [R] performance on AsNER.
+
+![](images/36e794983c1fc5e3c73affa2dd2285bc449320eecc9bbe5a9aa595dd5e1c16f1.jpg)
+
+AxomiyaBERTa by  $>5$  F1 points on AsNER, with an even greater improvement (10 F1 points) on WikiNER. AxomiyaBERTa also outperformed other baselines for both tasks, with the exception of MBERT on Wiki-based tasks. $^{9}$  Fig. 3 shows confusion matrices of performance on AsNER.
+
+IndicBERT and MBERT misclassified ORG tokens as LOC 16 times as much as AxomiyaBERTa. Specific cases include sub-tokens like निकागूर्थ ( /ni-ujɔːk/, “New York”) or निकागूर्थ ( /singapuːr/, “Singapore”), that are actually parts of entities like निकागूर्थ ( /e staɪ singapuːr/, “A-Star Singapore”) or निकागूर्थ रागूर्थ चन्देश निकागूर्थ ( /niujɔːk blad sentaːr/, “New York Blood Center”). This suggests that smaller, monolingual models like AxomiyaBERTa with a
+
+reduced sequence length and no NSP training objective are optimized for NE classification tasks with greater attention to local context (since the average sentence containing NEs is  $\sim 6$  tokens).
+
+Better overall performance on AsNER than on WikiNER can be partially attributed to having one fewer class and a more balanced distribution between categories. AsNER performance likely benefited from a greater phonological signal and more data to tune on (Table 3) whereas WikiNER text samples are, on average, longer than 128 tokens (AxomiyaBERTa's maximum token length) possibly causing a performance loss due to truncated context.
+
+Phonological Signals: A Disambiguation Tool Even though phonologically-aware AxomiyaBERTa took a hit on identifying  $O$  tokens, it compensated with improved results across other classes. Phonologically-aware AxomiyaBERTa also reduced misclassifications of  $ORG$  tokens as  $PER$  compared to all other models, including native AxomiyaBERTa. Specific cases include tokens that imply persons, e.g.,  $\text{रावलीतान्यत}$  or  $\text{िकार्त},$  but are actually part of  $ORG$  NEs, e.g.,  $\text{रावलीतान्यत}$ $\text{किकार्त्}$  ("Swaminathan Commission") or  $\text{िकार्त्}$ $\text{िकार्त्}$  "Saha Institute of Physics"). Similarly, in the WikiNER task, phonological attention reduced misclassification of  $B-ORG$  and  $I-ORG$  tokens as  $B-PER$  and  $I-PER$  respectively (see Appendix C). These results suggest phonological inputs help enrich embeddings of smaller-sized LMs to distinguish such ambiguous tokens.
+
+# 5.2 Long-context: Multiple Choice
+
+On Wiki-Titles, phonological AxomiyaBERTa does better with semantically harder multiple-choice sets, which have a higher average cosine similarity between the candidate options. Native AxomiyaBERTa fails on these samples. As shown
+
+in Table 5,  $\mathbf{P} + \mathbf{N}$  has the highest average cosine similarity between the sample sets, suggesting that there are cases where phonological signals compensate for low semantic variation among candidate options. On the other hand, native AxomiyaBERTa tends to do better with multiple-choice sets that have wider (relative) semantic variation within that set, on average. Since the overall distribution of embeddings in this task is still extremely close, this suggests that phonological signals are doing for Wiki-Titles what the embedding disperser did for Cloze-QA (see Sec. 3.2.1).
+
+<table><tr><td></td><td>P+N-</td><td>P-N+</td><td>P+N+</td><td>P-N-</td></tr><tr><td>Cos-sim</td><td>.98844</td><td>.98829</td><td>.98824</td><td>.98838</td></tr></table>
+
+Table 5: Average cosine similarities between within-set samples on the Wiki-Titles test set for native (N) and phonological (P) AxomiyaBERTa. “+” and “-” represent correct and incorrect samples respectively, e.g.,  $\mathbf{P} + \mathbf{N}$ - shows samples phonological AxomiyaBERTa answered correctly that the native variant did not.
+
+# 5.3 Novel Task: Event Coreference on Translated ECB+
+
+Table 6 shows event coreference resolution results on the translated ECB+ test set using a tuned affinity-threshold  $(T)$ . These results include both within- and cross-document system outputs from AxomiyaBERTa, other Transformer-based LMs, and a lemma-based heuristic.10
+
+AxomiyaBERTa often outperformed the lemmasimilarity baseline and other LMs. Native and phonological AxomiyaBERTa have the best MUC and BCUB F1 scores, respectively, while also outperforming all other Transformer-based LMs on BLANC and CoNLL F1. Phonologically-aware AxomiyaBERTa also outperforms native AxomiyaBERTa by almost 2 F1 points on CoNLL F1. More importantly, the phonological signals help detect more challenging coreferent links where mere surface-level lemma similarity can fail. While native and phonological AxomiyaBERTa performed comparably, the true positives retrieved by the phonological version contained a higher proportion of non-similar lemmas, which were usually missed by the lemma heuristic. Meanwhile, native AxomiyaBERTa retrieved results
+
+with more similar lemmas, labeling more nonsimilar lemma pairs as false negatives (Table 7). Compared to the other Transformer models, this also had the effect of increasing precision according to most metrics, though at the cost of decreasing recall. However, the increased precision was usually enough to increase F1 overall, pointing to the utility of phonological signals in detecting more challenging cases. We hypothesize that this is because these challenging pairs may consist of synonyms and/or loanwords, and phonological signals helped correlate these different surface forms, which in addition to the semantic information at the embedding level helps create coreference links.
+
+For instance,  $\text{क्वत्तालिकू}$  (/konsalti/, "consulting") and  $\text{रेमिकालिकू}$  (/indzinijai/, "engineering") denote two coreferent events pertaining to the same company (EYP Mission Critical Facilities). Since both are borrowed words that maintain the original phonological form, phonological signals can help pick out unique articulation beyond surface-level lemma similarity. Similarly, in cases of synonyms like  $\text{मुधान्यू}$  (/mittu/, "(of) death") and  $\text{मुधान्यू}$  ("killing"), which do not share surface-level similarity yet are coreferent, phonological signals can help. Where lemmas are already similar, phonological signals provide little extra information.
+
+We should note that for coreference, the specific metric used matters a lot. For instance, almost  $33\%$  of the  $\mathrm{ECB + }$  dataset across all three splits consists of singleton mentions. Since MUC score is not as sensitive to the presence of singletons as BCUB (Kubler and Zhekova, 2011), this could explain AxomiyaBERTa's (and XLM's) relative drop in performance on the BCUB metric. On the other hand, the lower CEAF-e F1 score may be due to CEAF-e's alignment algorithm, which tends to ignore correct coreference decisions when response entities are misaligned (Moosavi and Strube, 2016).
+
+Ablations between native and phonological AxomiyaBERTa showed that where lemmas for a pair of potentially coreferent events are identical (e.g., अयावर्ष - /aɪmb⁶o/, “start”), non-phonological representations primarily determine the pairwise scores and the coreference decision. Table 7 shows that even though phonological signals tend to disambiguate harder event pairs, decreased performance (e.g., MUC F1 phonological vs. native
+
+<table><tr><td rowspan="2">CDCR Models</td><td colspan="3">BCUB</td><td colspan="3">MUC</td><td colspan="3">CEAF-e</td><td colspan="3">BLANC</td><td>C-F1</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td></td></tr><tr><td>Lemma Baseline</td><td>75.81</td><td>60.24</td><td>67.14</td><td>64.59</td><td>54.25</td><td>58.97</td><td>61.36</td><td>73.25</td><td>66.78</td><td>74.97</td><td>60.40</td><td>64.66</td><td>64.29</td></tr><tr><td>XLM-100†</td><td>5.31</td><td>97.55</td><td>10.08</td><td>54.17</td><td>97.84</td><td>69.73</td><td>30.99</td><td>0.73</td><td>1.42</td><td>49.78</td><td>50.00</td><td>49.89</td><td>27.07</td></tr><tr><td>IndicBERT-BASE</td><td>74.48</td><td>51.93</td><td>61.19</td><td>44.03</td><td>21.94</td><td>29.29</td><td>40.80</td><td>65.59</td><td>50.31</td><td>52.09</td><td>55.41</td><td>52.93</td><td>46.93</td></tr><tr><td>MuRIL</td><td>93.53</td><td>48.33</td><td>63.73</td><td>68.18</td><td>9.23</td><td>16.26</td><td>41.56</td><td>85.09</td><td>55.85</td><td>54.78</td><td>53.31</td><td>53.91</td><td>45.28</td></tr><tr><td>AxomiyaBERTa</td><td>34.68</td><td>85.98</td><td>49.42</td><td>62.40</td><td>80.51</td><td>70.30</td><td>67.63</td><td>43.85</td><td>53.20</td><td>53.00</td><td>87.75</td><td>54.23</td><td>57.64</td></tr><tr><td>AxomiyaBERTa + Phon</td><td>70.00</td><td>64.58</td><td>67.18</td><td>64.11</td><td>44.71</td><td>52.68</td><td>50.18</td><td>68.57</td><td>50.18</td><td>56.22</td><td>68.65</td><td>59.19</td><td>59.27</td></tr></table>
+
+Table 6: Event coreference results on Assamese (translated) ECB+ test set from pairwise scorer using AxomiyaBERTa, compared with other Transformer-based LMs and the lemma-based heuristic. Bold indicates best overall performance per metric. "C-F1" is CoNLL F1.  ${}^{ \dagger  }$  We evaluate XLM-100 to compare performance on this task of a slightly larger model than XLM-R where most other major design factors remain the same.
+
+AxomiyaBERTa) could be due to native representations of the same-lemma pair being weakly correlated with the pairwise scores, a possibility when a coreferent event pair has high contextual dissimilarity. Phonological signals may add noise here.
+
+We also see that the lemma-based heuristic baseline is overall a very good performer. While this may be a property of the nature of coreference tasks in general or specific to a dataset (as a high percentage of coreferent events use the same lemma), we must also allow for the possibility that this may also be an artifact of translation noise. Since we used an automatically-translated version of the  $\mathrm{ECB + }$  corpus (albeit with some native speaker verification), and since Assamese is still a low-resource language, the decoder vocabulary of the translator may be limited, meaning that synonymous different-lemma pairs in the original corpus may well have been collapsed into same-lemma pairs in the translation, artificially raising the performance of the lemma heuristic.
+
+<table><tr><td>Models</td><td>TP</td><td>L1</td><td>L2</td><td>Diff-Rate</td></tr><tr><td>XLM-100</td><td>6,361</td><td>1,441</td><td>4,920</td><td>.773</td></tr><tr><td>IndicBERT</td><td>101</td><td>46</td><td>55</td><td>.545</td></tr><tr><td>MuRIL</td><td>62</td><td>21</td><td>41</td><td>.661</td></tr><tr><td>AxB</td><td>1,833</td><td>466</td><td>1,367</td><td>.746 (.98)</td></tr><tr><td>AxB + Phon</td><td>956</td><td>81</td><td>875</td><td>.915 (.93)</td></tr></table>
+
+Table 7: Distribution of same (L1) and different (L2) event lemma samples in the true positive (TP) distribution of the T-ECB+ test set. "Diff-Rate" is the percentage of different lemma samples within TPs  $(= L2 / TP)$ . Values in parentheses show the equivalent distribution within false negatives for comparison.
+
+# 6 Conclusion and Future Work
+
+In this paper, we presented a novel Transformer model for Assamese that optionally includes phonological signals. We evaluated on multiple tasks using novel training techniques and have
+
+demonstrated SOTA or comparable results, showing that phonological signals can be leveraged for greater performance and disambiguation for a low-resourced language. AxomiyaBERTa achieves SOTA performance on short-context tasks like AsNER and long-context tasks like Cloze-QA while also outperforming most other Transformer-based LMs on WikiNER, with additional improvement resulting from the phonologically-aware model. For challenging tasks like CDCR, we have shown that both AxomiyaBERTa outperformed other Transformer-based LMs on popular metrics like BCUB, MUC, and CoNLL F1.
+
+More generally, we have shown that strategic techniques for optimizing the embedding space and language-specific features like phonological information can lower the barrier to entry for training language models for LRLs, making it more feasible than before with lower amounts of data and a ceiling on compute power. Our experiments suggest phonological awareness boosts performance on many tasks in low-resource settings. Future models for other LRLs can leverage our ideas to train or fine-tune their own models. Since smaller models tend toward anisotropy, embedding dispersal may pave the way for more such performant LRL models.
+
+Future work may include incorporating phonological signals during pretraining instead of finetuning, carrying out evaluations against semantically harder tasks like paraphrasing or emotion detection, zero-shot transfer to similar languages, and a contrastive learning framework with a triplet loss objective for CDCR.
+
+Our trained checkpoints are available on HuggingFace at https://huggingface.co/Abhijnan/AxomiyaBERTa. We hope this resource will accelerate NLP research for encoding language-specific properties in LRLs.
+
+# Limitations
+
+Let us begin with the obvious limitation: AxiomiyaBERTa only works on Assamese. In addition, since Assamese comprises a number of dialects and we trained on internet-sourced data, we have no clear evidence regarding which dialects AxiomiyaBERTa is most suited to or if it performs as well on non-standard dialects.
+
+AxomiyaBERTa did not perform all that well on Wikipedia Title Selection, compared to other Transformer-based models. Our best result is on par with XLM-R and close to IndicBERT-BASE, but well below MBERT performance. We hypothesize that the amount of Wikipedia training data in MBERT is a cause of this, but we find that phonological attention makes a big difference in AxomiyaBERTa's performance (increasing accuracy from  $26\%$  to  $59\%$ ). Nonetheless, the reasons behind this subpar performance, and whether AxomiyaBERTa can be improved for this task without, say, overfitting to Wikipedia, need further investigation.
+
+# Ethics Statement
+
+Data Usage Because of the publicly-available, internet-sourced nature of our training data, we cannot definitively state that the current version of AxomiyaBERTa is free of bias, both in terms of outputs nor, as mentioned in the limitations section, if there are dialect-level biases toward or against certain varieties of Assamese that may be trained into the model. Such investigations are the topic of future research.
+
+# Resource Usage and Environmental Impact
+
+At 66M parameters, AxomiyaBERTa is a smaller language model that is relatively quick to train and run. Training was conducted on single GPU devices. Pretraining AxomiyaBERTa took approximately 3 days, and task-level fine-tuning took roughly 30 minutes for non-phonological AxomiyaBERTa and 1-2 hours for phonological AxomiyaBERTa (depending on the task). Training the pairwise scorer for CDCR took 12-19 minutes. Training and fine-tuning took place on the same hardware. For comparison, fine-tuning IndicBERT and MBERT on the AsNER dataset for evaluation took roughly 20-30 minutes each. These figures indicate that relative to work on other Transformer models, training and evaluating AxomiyaBERTa (including running other base
+
+lines for comparison) comes with a comparatively lower resource usage and concomitant environmental impact. This lower resource usage also has implications for the "democratization" of NLP, in that we have demonstrated ways to train a perform- mant model with fewer local resources, meaning less reliance on large infrastructures available to only the biggest corporations and universities.
+
+Human Subjects This research did not involve human subjects.
+
+# Acknowledgments
+
+We would like to thank the anonymous reviewers whose feedback helped improve the final copy of this paper. Special thanks to Ibrahim Khebour for helping with the phonological feature extraction process for the Wikipedia Section Title Prediction task.
+
+# References
+
+Shafiuddin Rehan Ahmed, Abhijnan Nath, James H. Martin, and Nikhil Krishnaswamy. 2023.  $2^{*}\mathfrak{n}$  is better than  $\mathfrak{n}^2$ : Decomposing Event Coreference Resolution into Two Tractable Problems. In *Findings of the Association for Computational Linguistics: ACL* 2023. ACL.  
+Emily Bender. 2019. The #benderrule: On naming the languages we study and why it matters. *The Gradient*, 14.  
+Himadri Bharali, Mayashree Mahanta, Shikhar Kumar Sarma, Utpal Saikia, and Dibyajyoti Sarmah. 2014. An analytical study of synonymy in Assamese language using WorldNet: Classification and structure. In Proceedings of the Seventh Global Wordnet Conference, pages 250-255.  
+Avi Caciularu, Arman Cohan, Iz Beltagy, Matthew Peters, Arie Cattan, and Ido Dagan. 2021. CDLM: Cross-document language modeling. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2648-2662, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Xingyu Cai, Jiaji Huang, Yuchen Bian, and Kenneth Church. 2021. Isotropy in the Contextual Embedding Space: Clusters and Manifolds. In International Conference on Learning Representations.  
+Oralie Cattan, Sophie Rosset, and Christophe Servan. 2021. On the cross-lingual transferability of multilingual prototypical models across NLU tasks. In Proceedings of the 1st Workshop on Meta Learning and Its Applications to Natural Language Processing, pages 36-43, Online. Association for Computational Linguistics.
+
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.  
+Agata Cybulska and Piek Vossen. 2014. Using a sledgehammer to crack a nut? lexical diversity and event coreference resolution. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 4545-4552, Reykjavik, Iceland. European Language Resources Association (ELRA).  
+David Demeter, Gregory Kimmel, and Doug Downey. 2020. Stolen probability: A structural weakness of neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2191-2197, 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.  
+Kawin Ethayarajh. 2019. How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. 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 55–65, Hong Kong, China. Association for Computational Linguistics.  
+Jun Gao, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tieyan Liu. 2018. Representation Degeneration Problem in Training Natural Language Generation Models. In International Conference on Learning Representations.  
+Barry Haddow and Faheem Kirefu. 2020. PMIndia-A Collection of Parallel Corpora of Languages of India. arXiv preprint arXiv:2001.09907.  
+Danesh Jain and George Cardona. 2004. The Indo-Aryan Languages. Routledge.  
+Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282-6293, Online. Association for Computational Linguistics.
+
+Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla, Gokul N.C., Avik Bhattacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. IndicNLPSuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for Indian languages. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4948-4961, Online. Association for Computational Linguistics.  
+Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja Nagipogu, Shachi Dave, Shruti Gupta, Subhash Chandra Bose Gali, Vish Subramanian, and Partha P. Talukdar. 2021. MuRIL: Multilingual Representations for Indian Languages. CoRR, abs/2103.10730.  
+Sandra Kübler and Desislava Zhekova. 2011. Singletons and coreference resolution evaluation. In Proceedings of the International Conference Recent Advances in Natural Language Processing 2011, pages 261-267.  
+Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.  
+Ritesh Kumar and Manas Jyoti Bora. 2018. Part-of-speech annotation of English-Assamese code-mixed texts: Two approaches. In Proceedings of the First International Workshop on Language Cognition and Computational Models, pages 94-103, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In International Conference on Learning Representations.  
+Sahinur Rahman Laskar, Abdullah Faiz Ur Rahman Khilji, Partha Pakray, and Sivaji Bandyopadhyay. 2020. EnAsCorp1.0: English-Assamese corpus. In Proceedings of the 3rd Workshop on Technologies for MT of Low Resource Languages, pages 62-68, Suzhou, China. Association for Computational Linguistics.  
+Jing Lu and Vincent Ng. 2018. Event coreference resolution: A survey of two decades of research. In *IJCAI*, pages 5479-5486.  
+Nafise Sadat Moosavi and Michael Strube. 2016. Which coreference evaluation metric do you trust? a proposal for a link-based entity aware metric. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 632-642, Berlin, Germany. Association for Computational Linguistics.
+
+David R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2018. Epitran: Precision G2P for many languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).  
+David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori Levin. 2016. PanPhon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475-3484, Osaka, Japan. The COLING 2016 Organizing Committee.  
+Abhijnan Nath, Rahul Ghosh, and Nikhil Krishnaswamy. 2022a. Phonetic, semantic, and articulatory features in Assamese-Bengali cognate detection. In Proceedings of the Ninth Workshop on NLP for Similar Languages, Varieties and Dialects, pages 41-53, Gyeongju, Republic of Korea. Association for Computational Linguistics.  
+Abhijnan Nath, Sina Mahdipour Saravani, Ibrahim Khebour, Sheikh Mannan, Zihui Li, and Nikhil Krishnaswamy. 2022b. A generalized method for automated multilingual loanword detection. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4996-5013, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Prachurya Nath, Prottay Kumar Adhikary, Pankaj Dadure, Partha Pakray, Riyanka Manna, and Sivaji Bandyopadhyay. 2022c. Image Caption Generation for Low-Resource Assamese Language. In Proceedings of the 34th Conference on Computational Linguistics and Speech Processing (ROCLING 2022), pages 263–272, Taipei, Taiwan. The Association for Computational Linguistics and Chinese Language Processing (ACLCLP).  
+Antoine Nzeyimana and Andre Niyongabo Rubungo. 2022. KinyaBERT: a morphology-aware Kinyarwanda language model. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5347-5363, Dublin, Ireland. Association for Computational Linguistics.  
+Kelechi Ogueji, Yuxin Zhu, and Jimmy Lin. 2021. Small data? no problem! exploring the viability of pretrained multilingual language models for low-resourced languages. In Proceedings of the 1st Workshop on Multilingual Representation Learning, pages 116-126, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1946-1958, Vancouver, Canada. Association for Computational Linguistics.
+
+Dhrubajyoti Pathak, Sukumar Nandi, and Priyankoo Sarmah. 2022. AsNER - annotated dataset and baseline for Assamese named entity recognition. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6571-6577, Marseille, France. European Language Resources Association.  
+Shruti Rijhwani, Jiateng Xie, Graham Neubig, and Jaime Carbonell. 2019. Zero-shot neural transfer for cross-lingual entity linking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6924-6931.  
+Shikhar Kr. Sarma, Himadri Bharali, Ambeswar Gogoi, Ratul Deka, and Anup Kr. Barman. 2012. A structured approach for building Assamese corpus: Insights, applications and challenges. In Proceedings of the 10th Workshop on Asian Language Resources, pages 21-28, Mumbai, India. The COLING 2012 Organizing Committee.  
+Shikhar Kumar Sarma and Rita Chakraborty. 2012. Structured and Logical Representations of Assamese Text for Question-Answering System. In Proceedings of the Workshop on Question Answering for Complex Domains, pages 27-38.  
+Jumi Sarmah, Shikhar Kumar Sarma, and Anup Kumar Barman. 2019. Development of Assamese rule based stemmer using WordNet. In Proceedings of the 10th Global WordNet Conference, pages 135-139.  
+Raphael Scheible, Fabian Thomczyk, Patric Tippmann, Victor Jaravine, and Martin Boeker. 2020. Gottbert: a pure German language model. arXiv preprint arXiv:2012.02110.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics.  
+Wei Shi and Vera Demberg. 2019. Next sentence prediction helps implicit discourse relation classification within and across domains. 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 5790-5796, Hong Kong, China. Association for Computational Linguistics.  
+Pedro Javier Ortiz Suárez, Benoit Sagot, and Laurent Romary. 2019. Asynchronous pipeline for processing huge corpora on medium to low resource infrastructures. In 7th Workshop on the Challenges in the Management of Large Corpora (CMLC-7). Leibniz-Institut für Deutsche Sprache.  
+Hasan Tanvir, Claudia Kittask, Sandra Eiche, and Kairit Sirts. 2021. EstBERT: A Pretrained
+
+Language-Specific BERT for Estonian. NoDaLiDa 2021, page 11.
+
+Antti Virtanen, Jenna Kanerva, Rami Ilo, Jouni Luoma, Juhani Luotolahti, Tapio Salakoski, Filip Ginter, and Sampo Pyysalo. 2019. Multilingual is not enough: BERT for Finnish. arXiv preprint arXiv:1912.07076.
+
+Jing Zhao, Yifan Wang, Junwei Bao, Youzheng Wu, and Xiaodong He. 2022. Fine- and coarse-granularity hybrid self-attention for efficient BERT. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4811–4820, Dublin, Ireland. Association for Computational Linguistics.
+
+# A Training Configuration
+
+Table 8 shows the pretraining configuration for AxomiyaBERTa.
+
+# B Further Details on Embedding Disperser
+
+Fig. 4 shows KDE plots for outputs of different components of the embedding disperser, showing the contrast between features within-set and beyond-set for Cloze-QA samples, and showing the difference between AxomiyaBERTa with phonological awareness and without. The option_cos label (brown) shows an interesting phenomenon. This is the output of the embedding disperser at inference (Auxiliary Discriminator in Fig. 2) and represents a 128-dimensional embedding output from the [CLS] token concatenated with arg2 or the candidate answer input. We see a distinct shift in cosine similarity scores between within-set and beyond-set with one peak very close to 1 in the case of the within-set pairs while getting clearly dispersed to a lower cosine similarity score in the case of beyond-set pairs. This phenomenon is even further accentuated by feeding phonological signals to the disperser. In this case, as shown in the top right plot, the cosine similarity peak for option.cos has a much higher density compared to the non-phonological disperser while the overall distribution is shifted to a higher cosine similarity.
+
+Another interesting trend is the linear_sigmoid label (red) which is the sigmoidal output of the linear layer of the disperser, trained with a combination of cosine embedding loss and BCE loss when fed an input of the combined arg1 and arg2 representations generated with the special trigger tokens. In this
+
+case, feeding phonological signals to the model reduces dispersion (an inverse trend) in the cosine similarities between within-set and beyond-set pairs (as seen in the top-left plot where this label has a narrower top with a wider bottom). However, this reverse effect is less pronounced than that seen in the option_cos cosine similarity plot, perhaps due to richer contextual information carried by the trigger token representations (the inputs to this layer). In other words, and as shown in the arg_cos_sim plot, its dispersion between the within- and beyond-set pairs suggests why such an effect is less-pronounced.
+
+Works such as Cai et al. (2021) present evidence of such global token anisotropy in other BERT and GPT-model variants while also suggesting ways to locate/create local isotropic spaces more susceptible for NLP tasks. Interestingly, cosine similarities of output embeddings from our Auxiliary Discriminator (option_cos in Fig. 1) show a marked difference in the extent of anisotropy between within-set and beyond-set pairs, a phenomenon further accentuated with additional phonological signals (top right plot in Fig. 4). These experiments suggest that a combination of our embedding disperser architecture together with phonological signals (Sec. 3.2.2 for more details) can effect a shift towards local spaces of isotropy in the embedding space of the fine-tuned AxomiyaBERTa model for Cloze-QA and potentially other tasks.
+
+# C Further Discussion on Short-Context Results
+
+Fig. 5 shows native and phonological AxomiyaBERTa performance on WikiNER. We see comparative performance, but with phonological signals there are fewer confusions of  $B-ORG$  with  $B-PER$  and  $I-ORG$  with  $I-PER$ . Specific examples are similar to those seen in Sec. 5.1, e.g., "रागल्तार्वत (रागल्तार्वत)" ("Swaminathan [Commission]") or "सार्या (सिन्या अधिक शिलयो कृपि सिन्या)" ("Saha [Institute of Physics]"). Being organizations named after people, this is a case where phonological signals actually help. Interestingly, phonological signals also help with NER even when the NEs are broken down into BIO chunks, which was not the case in AsNER. We should observe that with phonological signals, there is an increase in  $B-LOC$  tokens classified as  $B-PER$  tokens, which is the topic of future investigation.
+
+![](images/5578b745139719e26530edd34fe0aaa65af9e2eec37e4e6c7c57344930da282f.jpg)
+
+![](images/572dbbee3d75b008905b991ceaf8aec5801a25caf1e07fcd5ba0093162d40a22.jpg)
+
+![](images/bcb6b3b86742109e70c98cefd6cd5c17b559ff6ed3452e12fc1fec2dd4738afa.jpg)  
+Figure 4: Kernel density estimation plots from various feature sets of the embedding disperser model. The figure on the top left represents "within set": cosine similarities between each set of candidate answers plus context and the remaining three pairs in that set, e.g., if  $Q$  is the question/context and  $A / B / C / D$  are the candidate answers,  $S_{C}(Q + A, Q + i)$ , where  $i$  represents one of the remaining three candidates  $B, C, D$ . The figure on the bottom left represents "beyond-set" cosine similarities: all pairs in a candidate-plus-answers set are compared to other such sets with the cosine similarity metric. We run our experiments for 100 randomly selected sets from the Cloze-QA dev set. The top right (within-set) and bottom right (beyond-set) figures are equivalent figures for our models with phonological awareness.
+
+![](images/f6bb3ce2db93eb8380e0ea1da34790f4c2a815d136d9dbc680a2ad1d9fae7489.jpg)
+
+<table><tr><td>Parameters</td><td>Config</td></tr><tr><td>architecture</td><td>AlbertForMaskedLM</td></tr><tr><td>attention_probs_dropoutprob</td><td>0.1</td></tr><tr><td>bos_token_id</td><td>2</td></tr><tr><td>classifier_dropoutprob</td><td>0.1</td></tr><tr><td>embedding_size</td><td>128</td></tr><tr><td>eos_token_id</td><td>3</td></tr><tr><td>hidden_ACT</td><td>gelu</td></tr><tr><td>hidden_dropoutprob</td><td>0.1</td></tr><tr><td>hidden_size</td><td>768</td></tr><tr><td>initializer_range</td><td>0.02</td></tr><tr><td>inner_group_num</td><td>1</td></tr><tr><td>intermediate_size</td><td>3072</td></tr><tr><td>layer_normnano</td><td>1e-05</td></tr><tr><td>max_position_embeddings</td><td>514</td></tr><tr><td>num attentive_heads</td><td>12</td></tr><tr><td>num_hidden_groups</td><td>1</td></tr><tr><td>num_hidden_layers</td><td>6</td></tr><tr><td>position_embedding_type</td><td>“absolute”</td></tr><tr><td>transformers_version</td><td>“4.18.0”</td></tr><tr><td>vocab_size</td><td>32001</td></tr></table>
+
+Table 8: AxomiyaBERTa Model configuration trained on a monolingual Assamese corpus.
+
+![](images/33adaceb5f8e6f8e94062e3c2b55ea9f6bd278a176855425eb196e7aef805ba3.jpg)  
+Figure 5: Confusion matrices showing AxomiyaBERTa performance on WikiNER without [L] and with [R] phonological awareness.
+
+![](images/cc9b13ce7e24184a179cd856ee44b0e36d191fcc23f757450597839d6f5cf080.jpg)
+
+# D Further Discussion on Pairwise Scorer for CDCR on Assamese ECB+
+
+The lemma-based heuristic comes from the fact that a large proportion of coreferent mention pairs can be identified simply because they use the same lemma. These "easy" cases give coreference a very high baseline even when this naive heuristic is used. The long tail of "harder" pairs require more sophisticated approaches (Ahmed et al., 2023).
+
+Fig. 6 shows the affinity scores from the pairwise scorer using various model outputs. AxomiyaBERTa is shown in the top left, followed by (left-to-right, top-to-bottom) XLM-100, MuRIL, and IndicBERT. We see that AxomiyaBERTa clearly has a more defined separation between the labels, with positive/coreferent samples having
+
+higher affinity scores (accounting for the imbalanced distribution of coreferent vs. non-coreferent pairs) compared to the other models. In particular, XLM-100 shows almost identical ranges of scores for coreferent and non-coreferent pairs, with the only significant difference being the number of each kind of sample, which results in the spike around  $T = -1.94$  (cf. Sec. 3.2.6).
+
+![](images/05bb066e454aacc0cd1a76ea0bd25bfbc2177183686a10e3c9d507e947ffbbd3.jpg)
+
+![](images/2313d64ea1df619030ec92e6d071e1e609f2a8e7dc6a5f820e48b0887dc81ef7.jpg)
+
+![](images/ab5115846402f459822d1ec93840a460f14c813bad42be047a824b8740a4c3b0.jpg)  
+Figure 6: Kernel density estimation plots of affinity scores from the pairwise scorer for native AxomiyaBERTa compared to baselines from other Transformer-based LMs.
+
+![](images/73d11d11155a8f26a4a4ee6368c0665bd34891607111f7b45279ee8dff4a1aab.jpg)
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Required limitations section after conclusion
+
+A2. Did you discuss any potential risks of your work?
+
+Ethics section (end of paper)
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Abstract and introduction
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+The paper results in a new language model, datasets used in training are discussed in Section 3.1. Other artifacts (e.g., preprocessing packages) are discussed in Section 3 and subsections
+
+B1. Did you cite the creators of artifacts you used?
+
+Section 3.*
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Section 6. AxomiyaBERTa will be made freely available upon publication.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Section 6. AxomiyaBERTa will be made freely available for use upon publication. Since the artifacts we used in training and evaluation are popularly used in the AI/NLP/CL domain (both academia and industry) with proper citations, we have taken all measures to ensure that our usage is consistent with their intended use.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+We use publicly-available datasets like and discuss the risks in the ethics statement.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Section 1, and limitations and ethics section
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Section 4
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+# C Did you run computational experiments?
+
+Section 4
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Section 3, Section 4
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Section 3 and subsections on fine-tuning
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+Section 5
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Section 3. These details are reported where available but some packages used (e.g., PanPhon) do not have different versions available.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/AxomiyaBERTa_ A Phonologically-aware Transformer Model for Assamese/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad42da8686c5fea47185b38bcfdbc980c85954c5151aba1bd650922974581275
+size 522719

2023/B2T Connection_ Serving Stability and Performance in Deep Transformers/304a7eb6-4126-4acb-9ec0-8da7fd06045b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f33dee1e084e01613347c76b2247adc04bfc463fb3a0924751659eb1f190f2b0
+size 803868

2023/B2T Connection_ Serving Stability and Performance in Deep Transformers/full.md

@@ -0,0 +1,524 @@
+# B2T Connection: Serving Stability and Performance in Deep Transformers
+
+Sho Takase†* Shun Kiyono† Sosuke Kobayashi‡ Jun Suzuki‡  
+†LINE Corporation ‡Tohoku University  
+{sho.takase, shun.kiyono}@linecorp.com  
+sosk@preferred.jp  
+jun.suzuki@tohoku.ac.jp
+
+# Abstract
+
+From the perspective of the layer normalization (LN) positions, the architectures of Transformers can be categorized into two types: Post-LN and Pre-LN. Recent Transformers tend to be Pre-LN because, in Post-LN with deep Transformers (e.g., those with ten or more layers), the training is often unstable, resulting in useless models. However, Post-LN has consistently achieved better performance than Pre-LN in relatively shallow Transformers (e.g., those with six or fewer layers). This study first investigates the reason for these discrepant observations empirically and theoretically and made the following discoveries: 1, the LN in Post-LN is the main source of the vanishing gradient problem that leads to unstable training, whereas Pre-LN prevents it, and 2, Post-LN tends to preserve larger gradient norms in higher layers during the back-propagation, which may lead to effective training. Exploiting the new findings, we propose a method that can provide both high stability and effective training by a simple modification of Post-LN. We conduct experiments on a wide range of text generation tasks. The experimental results demonstrate that our method outperforms Pre-LN, and enables stable training regardless of the shallow or deep layer settings. Our code is publicly available at https://github.com/takase/b2t_Connection.
+
+# 1 Introduction
+
+To prevent the vanishing (or exploding) gradient problem in the training of a deep neural network (DNN), various techniques, such as batch normalization (Ioffe and Szegedy, 2015) and residual connection (Srivastava et al., 2015; He et al., 2016a), have been proposed and widely used in almost all recent DNNs. Transformer (Vaswani et al., 2017) employs the layer normalization (Ba et al., 2016) for this purpose. Transformer is currently the most successful model architecture
+
+![](images/eaaa224b18e85f74e1a38e1a2f30afc3bc0cc45231ba2096cf8bae2913417e88.jpg)  
+(a) Train loss  
+Figure 1: Loss values of 18 layered Transformer-based encoder-decoder architectures.
+
+![](images/2645a67aa2b6b33d79dcc8bb4817f817e6b9feffa7582c568abac5d36df3d216.jpg)  
+(b) Valid loss
+
+in DNNs. It was firstly developed for applying sequence-to-sequence tasks, such as machine translation (Vaswani et al., 2017), summarization (Takase and Okazaki, 2019), and automatic speech recognition (ASR) (Wang et al., 2020), and is currently used in speech, vision, and many other information processing research fields.
+
+As reported in the batch normalization literature (He et al., 2016b), the position of the normalization layers primarily affects both the stability and resultant performance of a trained model. In Transformers, some previous studies have investigated the impact of the layer normalization positions (Wang et al., 2019; Xiong et al., 2020). There are currently two major layer normalization positions in Transformers: Pre-Layer Normalization (Pre-LN) and Post-Layer Normalization (PostLN). Pre-LN applies the layer normalization to an input for each sub-layer, and Post-LN places the layer normalization after each residual connection. The original Transformer (Vaswani et al., 2017) employs Post-LN. However, many recent studies have suggested using Pre-LN (Wang et al., 2019; Baevski and Auli, 2019; Brown et al., 2020) because the training of deep Transformers (e.g., those with ten or more layers) using Post-LN is often unstable, resulting in useless models. Figure 1
+
+shows loss curves for an actual example; the training of 18 layered Transformer encoder-decoders (18L-18L) on a widely used WMT English-to-German machine translation dataset. These figures clearly show that the Post-LN Transformer encoder-decoders fail to train the model. However, in contrast, Liu et al. (2020) reported that Post-LN consistently achieved better performance than Pre-LN on a machine translation task when they used 6 layered (relatively shallow, 6L-6L) Transformers.
+
+This paper focuses specifically on such discrepancies between Pre-LN and Post-LN in configurations with various number of layers. We investigate the sources of the instability of training in deep configurations and the superior performance in shallow configurations for Post-LN, compared with that for Pre-LN, to understand the essentials of the differences between Pre-LN and Post-LN. We discover that the layer normalization in Post-LN is the main source of the vanishing gradient problem that leads to unstable training, whereas Pre-LN prevents it, as shown in Figure 1. In particular, we clarify that the layer normalization is a significant factor of the vanishing gradient problem by comparing the input/output vector norms of gradient flows for each layer normalization during back-propagation. These analyses bring us a novel idea that can satisfy higher stability by skipping over layer normalizations and provide better performance than Pre-LN regardless of their layer sizes. Consequently, we propose a method that is based on Post-LN Transformers but has additional residual connections to enable stable training.
+
+We conduct experiments on a wide range of text generation tasks, namely machine translation, summarization, language modeling, and ASR. The experimental results lead to the following three new major findings:
+
+1. Post-LN Transformers achieve better performance than Pre-LN Transformers on text generation tasks (not only machine translation (Liu et al., 2020) but also other tasks). Thus, Post-LN is superior to Pre-LN if the problem of its unstable training can be solved.  
+2. Our modification enables Post-LN Transformers to stack many layers.  
+3. Our method can maintain the performance advantage of Post-LN and mitigate its unstable training property, thus providing better performance than Pre-LN.
+
+# 2 Post-LN and Pre-LN Transformers
+
+We briefly describe Post-LN and Pre-LN Transformers in this section. The original Transformer (Vaswani et al., 2017) uses Post-LN, in which layer normalizations are located after each residual connection. Let  $x$  be an input of a sublayer, and  $\mathcal{F}(\cdot)$  be a sub-layer of a Transformer, such as a feed-forward network or multi-head attention. Post-LN is defined as follows:
+
+$$
+\operatorname {P o s t L N} (x) = \operatorname {L N} (x + \mathcal {F} (x)), \tag {1}
+$$
+
+where  $\mathrm{LN}(\cdot)$  is the layer normalization function.
+
+In contrast, Pre-LN places the layer normalization before an input of each sub-layer:
+
+$$
+\operatorname {P r e L N} (x) = x + \mathcal {F} (\operatorname {L N} (x)). \tag {2}
+$$
+
+Figure 2 (a) and (b) illustrate Post-LN and Pre-LN Transformer architectures, respectively.
+
+# 3 Gradients of Transformer Layers
+
+As described in Liu et al. (2020), the vanishing gradient problem often occurs in Post-LN Transformers. Figure 3 shows the gradient norms of each layer for the (a) encoder-side and (b) decoder-side at the beginning of training, when 18L-18L Transformer encoder-decoders are trained on a widely used machine translation dataset (the WMT English-to-German dataset). Focus on the decoder-side of Post-LN as illustrated in Figure 3 (b). This figure shows that shallower layers have smaller gradient norms. In other words, the vanishing gradient occurs in the decoder-side of Post-LN because its gradient norms exponentially decay as they are back-propagated to shallower layers. This result is consistent with the previous study (Liu et al., 2020). We consider that this vanishing gradient causes the difficulty of stacking many layers with the Post-LN setting, as shown in Figure 1.
+
+To investigate the vanishing gradient empirically in more detail, we measure the gradient norms of parts (1) - (5) of Figure 2 (a). Figure 4 shows the gradient norms of each part in the 18th layer<sup>1</sup>. This figure shows that the gradient norms decrease drastically from (4) to (3) and (2) to (1). These parts correspond to layer normalizations, as shown in Figure 4. This suggests that layer normalizations in Post-LN Transformers are probably the cause of the vanishing gradient problem.
+
+![](images/15e081236e829149b336a1550bc65eb3c4caaaa0be7ee92d62d4ccf5dd7533fc.jpg)  
+(a) Post-LN
+
+![](images/cbf418b7b8681a696c5f61e111e62eabf4ffcd4a99e9e8cb174a1cf2f9a1559e.jpg)  
+(b) Pre-LN
+
+![](images/0f365fd87ce4366b3ac4c20bc02df3868b7defbc58708c74e2a2b85b5e3a2d69.jpg)  
+(c) Post-LN with B2T connection  
+Figure 2: Transformer-based encoder-decoder architectures for (a) Post-LN, (b) Pre-LN, and (c) Post-LN with our proposed B2T connection.
+
+![](images/c1ad4392d8334c00fffb187dbde919488073a4488228d8a61936a02cd4867fcb.jpg)  
+(a) Encoder side
+
+![](images/9a479426a0a23a207bdc6482f3b3fb02d7589a16eca5f5220624c2f9fc876d59.jpg)  
+(b) Decoder side
+
+To investigate the difference between the gradient flows of Post-LN and those of Pre-LN theoretically, we calculate the derivatives of Equations (1) and (2), as follows:
+
+$$
+\frac {\partial \operatorname {P o s t L N} (x)}{\partial x} = \frac {\partial \operatorname {L N} (x + \mathcal {F} (x))}{\partial (x + \mathcal {F} (x))} \left(I + \frac {\partial \mathcal {F} (x)}{\partial x}\right), \tag {3}
+$$
+
+$$
+\frac {\partial \operatorname {P r e L N} (x)}{\partial x} = I + \frac {\partial \mathcal {F} (\mathrm {L N} (x))}{\partial \mathrm {L N} (x)} \frac {\partial \mathrm {L N} (x)}{\partial x}, \tag {4}
+$$
+
+where  $I$  is the identity matrix. As Equation (3), the derivative of Post-LN is equal to the product of two derivatives: one is the layer normalization, and the other consists of the residual connection and sub-layer  $\mathcal{F}$ . In contrast, in Pre-LN, the derivative of the residual connection is isolated from the term related to the derivative of the layer normalization.
+
+![](images/0b4655451db07243b417ec52b72a03b403a1aa9fe0931fa19e581b152b80df8f.jpg)  
+Figure 3: Gradient norms of 18 layered Transformer-based encoder-decoder architectures.
+
+![](images/6893798d7338f995ef00220c71fda66f8f7ac1c7ecefb62c285f79ea43d70af6.jpg)  
+Figure 4: Gradient norms of each location in the 18th decoder for the 18 layered Post-LN Transformer encoder-decoder on WMT English-to-German translation training data.
+
+The difference between these equations implies that the residual connection in Pre-LN prevents the vanishing gradient because it retains the gradients of upper layers even if the derivative of the layer normalization decreases gradients drastically.
+
+# 4 Transformations by Each Layer
+
+As described, it is difficult to stack many layers in Post-LN Transformers because the vanishing gra
+
+![](images/d84aed36075a16cb95c31e68cc59cf5c70652ac63e34ff6be27d49b76e55ad95.jpg)
+
+![](images/5453b285a5d5ded6046bd1097e35be822dad2b3cea2e038669012fd2b133d928.jpg)
+
+![](images/92c57287f673ef44fd16a32f839e627b176fdf05d5a5afcddef7051c5a730d0f.jpg)
+
+![](images/e7bbe43fc4cbffc617530b5cd7f9f144672d4e400b21e37a59e3b72dc050e80b.jpg)  
+(a) Post-LN
+
+![](images/466cc0fd8a27420c65eb2aca6af3a8d3447fe92fb8b6073f5c3395aa9964c138.jpg)  
+(b) Pre-LN
+
+![](images/c15bbe3943e935f6d27126aa7ee27069dcf0525ccb2d3990cb9f574a99052279.jpg)  
+(c) B2T connection  
+Figure 5: Cosine similarities between the outputs of each pair of layers.
+
+dient problem occurs. Although Pre-LN is more stable in training, Post-LN can achieve better performance if training succeeds (see Section 6). In this section, we explore the reason for this difference in performance.
+
+Focus Pre-LN in Figure 3. In contrast to PostLN, in Pre-LN, a deeper (higher) layer has a smaller gradient norm. Thus, the parameters of higher layers are not required to change dramatically from their initial values. This implies that higher layers in Pre-LN are not sufficiently effective.
+
+To investigate the effectiveness of higher layers, we focus on the transformations by each layer. Figure 5 shows the average cosine similarities between the outputs of each pair of layers for 6L-6L Transformer encoder-decoders trained on the WMT dataset when several sequences are input. This figure indicates that the lower-left similarities of Pre-LN are higher than those of Post-LN. This result means that the outputs of shallow layers are similar to the output of the final layer in Pre-LN, but not in Post-LN. Consequently, higher layers in Pre-LN are less effective than those in Post-LN if training succeeds.
+
+We consider that the residual connection in Pre-LN causes this phenomenon. As Equation (2) shows, in Pre-LN, an input  $x$  skips over the sublayer  $\mathcal{F}(\cdot)$  by the residual connection. Thus, the input  $x$  is directly connected to the final layer output. This property makes the training stable, as described in Section 3, but causes high similarities between the outputs of the various layers. Therefore, we consider that Pre-LN underperforms Post-LN because the residual connection in Pre-LN reduces the effectiveness of its higher layers. In contrast, in Post-LN, larger gradient norms in higher layers (as shown in Figure 3) make higher layers more ef
+
+fective (as shown in Figure 5) but it is necessary to prevent the vanishing gradient problem in shallow layers when we stack many layers.
+
+# 5 Modification for Stable Training in Post-LN: Bottom-to-Top Connection
+
+This section introduces a modification that makes the training of Post-LN more stable while preserving its high performance. This modification comprises an additional residual connection to mitigate the vanishing gradient in Post-LN by enabling many layers to be stacked.
+
+As discussed in the previous sections, we need a term that retains gradients in the derivatives, as in Equation (4), to prevent the vanishing gradient. To satisfy this requirement, we propose a residual connection that skips over all layer normalizations except the final one in each layer. Our introduced connection ties an input of a layer to the result of the feed-forward network (FFN), as illustrated by the red arrows in Figure 2 (c). We call this connection Bottom-to-Top (B2T) connection, which is formalized in the following equation:
+
+$$
+x _ {i n p} + x _ {f f n} + \operatorname {F F N} \left(x _ {f f n}\right), \tag {5}
+$$
+
+where  $x_{inp}$  is an input of a layer,  $\mathrm{FFN}(\cdot)$  is an FFN, and  $x_{ffn}$  is an input of the FFN. In short,  $x_{inp}$  skips the layer normalizations after the self-attention and encoder-decoder cross-attention. Because the derivative of  $x_{inp}$  is isolated from the terms related to the derivatives of the layer normalizations just behind the attention sub-layers, it retains gradients, as in Pre-LN. For example, in an encoder-side,  $x_{ffn}$  is as follows:
+
+$$
+x _ {f f n} = \operatorname {L N} (\operatorname {S e l f A t t n} (x _ {i n p}) + x _ {i n p}), \tag {6}
+$$
+
+where SelfAttn(·) is a self-attention network. Thus, Equation (5) can be written as follows:
+
+$$
+\begin{array}{l} x _ {i n p} + \mathrm {L N} (\mathrm {S e l f A t t n} (x _ {i n p}) + x _ {i n p}) \\ + \operatorname {F F N} (\operatorname {L N} (\operatorname {S e l f A t t n} (x _ {i n p}) + x _ {i n p})), \tag {7} \\ \end{array}
+$$
+
+The derivative of this equation is the following equation:
+
+$$
+\begin{array}{l} I + \frac {\partial (\mathrm {L N} (\operatorname {S e l f A t t n} (x _ {i n p}) + x _ {i n p}))}{\partial x _ {i n p}} \\ + \frac {\partial (\mathrm {F F N} (\mathrm {L N} (\text {S e l f A t t n} (x _ {i n p}) + x _ {i n p})))}{\partial x _ {i n p}}, \tag {8} \\ \end{array}
+$$
+
+Because this derivative contains  $I$ , which is unrelated to the derivatives of internal layer normalizations, our B2T connection (i.e.,  $x_{inp}$ ) helps to propagate gradients. For a decoder-side, we can prove this property in the same manner.
+
+Figure 3 (b) indicates that B2T connection mitigates the vanishing gradient of 18L-18L encoder-decoders. Moreover, we locate B2T connection before the final layer normalization in each layer to avoid a direct connection to the final layer output based on the discussion in Section 4. Thus, B2T connection preserves the property of Post-LN with respect to the transformations performed by each layer, as illustrated in Figure 5 (c) $^2$ .
+
+# 6 Experiments
+
+Through experiments, we indicate following three findings.
+
+- Post-LN Transformers achieve better performance than Pre-LN Transformers if their training succeeds.  
+- B2T connection enables the training of deep Transformers with the Post-LN configuration.  
+- Our modification preserves the performance advantage of Post-LN Transformers, which therefore outperform Pre-LN Transformers.
+
+We describe the essential experimental configurations in this section. Appendix A presents more details, such as the hyper-parameters and computational budgets.
+
+# 6.1 Machine Translation
+
+# 6.1.1 Dataset
+
+The machine translation task has been widely used to investigate the performance of Transformer-based methods since the original Transformer (Vaswani et al., 2017; Ott et al., 2018; Wang et al., 2019; Xiong et al., 2020; Liu et al., 2020). We adopted the widely used WMT English-to-German training dataset (Vaswani et al., 2017; Ott et al., 2018), which contains 4.5M
+
+sentence pairs. We applied the byte-pair-encoding (BPE) algorithm (Sennrich et al., 2016) to construct a vocabulary set in the same manner as previous studies. We set the number of BPE merge operations to  $32\mathrm{K}$  and shared the vocabulary between the source and target languages. We used newstest2010-2016 to investigate the performance, following Takase and Kiyono (2021).
+
+# 6.1.2 Methods
+
+We compare Post-LN, Pre-LN, and Post-LN with our B2T connection (B2T connection) Transformers. We used  $\text{fairseq}^3$  (Ott et al., 2019) as an implementation of Transformers. We stacked 6 and 18 layers for the encoders and decoders (6L-6L and 18L-18L) as the widely used configuration and deep configuration, respectively. We used the Transformer (base) setting for dimension sizes of internal layers. In addition to the above methods, we evaluate the following five methods, which are recent approaches that enable the training of deep Transformers. We used the same hyper-parameters for all methods except T-Fixup. For T-Fixup, we used the hyper-parameters reported in Huang et al. (2020) to prevent divergence.
+
+DLCL To make Transformers deep, Wang et al. (2019) proposed dynamic linear combination of layers (DLCL), which uses the weighted sum of the lower layers as an input of a layer. In contrast to our B2T connection, which is an additional connection within each layer, DLCL uses a connection among layers. We apply DLCL to Post-LN Transformers. We used the official implementation<sup>4</sup>.
+
+Admin Liu et al. (2020) proposed adaptive model initialization (Admin), which uses additional parameters to stabilize the training of Post-LN Transformers. This method requires the variances of internal layers to initialize the additional parameters. Thus, this method first processes several forward steps for the initialization, and then conducts the actual training. In a nutshell, this method incurs additional computational costs. We used the official implementation<sup>5</sup>.
+
+T-Fixup Huang et al. (2020) proposed an initialization scheme for Transformers, T-Fixup, to perform stable training without the learning rate warm-up and layer normalizations. Because this method can remove the cause of the vanishing gradient, we can
+
+<table><tr><td>Method</td><td>2010</td><td>2011</td><td>2012</td><td>2013</td><td>2014</td><td>2015</td><td>2016</td><td>Average</td></tr><tr><td colspan="9">Enc-Dec: 6L-6L</td></tr><tr><td>Post-LN</td><td>24.27</td><td>22.06</td><td>22.43</td><td>26.11</td><td>27.13</td><td>29.70</td><td>34.40</td><td>26.59</td></tr><tr><td>Pre-LN</td><td>24.03</td><td>21.77</td><td>22.08</td><td>25.63</td><td>26.27</td><td>29.07</td><td>33.84</td><td>26.10</td></tr><tr><td>DLCL (Wang et al., 2019)</td><td>23.94</td><td>22.00</td><td>22.24</td><td>26.11</td><td>27.37</td><td>29.71</td><td>34.26</td><td>26.52</td></tr><tr><td>Admin (Liu et al., 2020)</td><td>24.32</td><td>21.79</td><td>22.17</td><td>26.26</td><td>27.14</td><td>29.61</td><td>34.12</td><td>26.49</td></tr><tr><td>T-Fixup (Huang et al., 2020)</td><td>24.09</td><td>21.98</td><td>22.04</td><td>25.96</td><td>26.92</td><td>29.45</td><td>34.56</td><td>26.43</td></tr><tr><td>RealFormer (He et al., 2021)</td><td>24.18</td><td>22.02</td><td>22.17</td><td>26.02</td><td>26.98</td><td>29.36</td><td>34.15</td><td>26.41</td></tr><tr><td>DeepNet (Wang et al., 2022)</td><td>24.08</td><td>21.76</td><td>22.09</td><td>25.90</td><td>26.85</td><td>29.62</td><td>34.39</td><td>26.38</td></tr><tr><td>B2T connection</td><td>24.12</td><td>21.93</td><td>22.29</td><td>26.31</td><td>26.84</td><td>29.48</td><td>34.73</td><td>26.53</td></tr></table>
+
+<table><tr><td colspan="9">Enc-Dec: 18L-18L</td></tr><tr><td>Post-LN</td><td colspan="7">Training failed (See Figure 1)</td><td>N/A</td></tr><tr><td>Pre-LN</td><td>24.07</td><td>21.98</td><td>22.40</td><td>26.28</td><td>27.36</td><td>29.74</td><td>34.16</td><td>26.57</td></tr><tr><td>DLCL (Wang et al., 2019)</td><td>24.20</td><td>22.51</td><td>22.83</td><td>26.59</td><td>27.97</td><td>30.24</td><td>33.98</td><td>26.90</td></tr><tr><td>Admin (Liu et al., 2020)</td><td>24.56</td><td>22.17</td><td>22.62</td><td>26.48</td><td>27.99</td><td>30.35</td><td>33.88</td><td>26.86</td></tr><tr><td>T-Fixup (Huang et al., 2020)</td><td>24.45</td><td>22.29</td><td>22.76</td><td>26.57</td><td>27.71</td><td>30.13</td><td>34.69</td><td>26.94</td></tr><tr><td>RealFormer (He et al., 2021)</td><td>24.32</td><td>22.42</td><td>22.68</td><td>26.59</td><td>28.58</td><td>30.36</td><td>33.71</td><td>26.95</td></tr><tr><td>DeepNet (Wang et al., 2022)</td><td>24.70</td><td>22.40</td><td>22.92</td><td>26.85</td><td>28.21</td><td>30.60</td><td>34.25</td><td>27.13</td></tr><tr><td>B2T connection</td><td>24.62</td><td>22.51</td><td>22.86</td><td>26.74</td><td>28.48</td><td>30.99</td><td>34.93</td><td>27.30</td></tr></table>
+
+Table 1: BLEU scores of each method on WMT newstest2010-2016 and their averages.
+
+stack many layers. We used the official implementation<sup>6</sup>.
+
+RealFormer To improve the performance of Transformers, He et al. (2021) proposed RealFormer, which introduces additional connections into attention sub-layers. Although their motivation is not addressing the vanishing gradient problem, their method is similar to ours with respect to the use of additional connections.
+
+DeepNet Wang et al. (2022) proposed DeepNorm, which uses a weight that corresponds to the number of layers in a residual connection before layer normalizations to stabilize Post-LN based Transformers. They also provided the combination of the initialization scheme and DeepNorm as DeepNet.
+
+# 6.1.3 Results
+
+We measured case-sensitive detokenized BLEU scores with SacreBLEU (Post, 2018). Table 1 shows BLEU scores of each method on newstest2010-2016 and the averaged scores of them. Since the BLEU score is precision-based n-gram overlapping between the model output and correct examples, a higher score represents better performance.
+
+![](images/3eb1e6ad9a2703dfa266754c794e4e9e70a6932623135c2b2989db195bb0a3ff.jpg)  
+Figure 6: Negative Log-Likelihood (NLL) on validation data (newstest2013) when we stack 18 layers.
+
+The upper part of Table 1 shows results in the 6L-6L configuration. This part indicates that Post-LN achieved better scores than Pre-LN on all test sets. In addition, B2T connection outperformed Pre-LN on all test sets. Thus, these methods are superior to Pre-LN when the total number of layers is small.
+
+The lower part of Table 1 shows results in the 18L-18L configuration. This part shows that the training of Post-LN failed, and thus we cannot successfully stack 18L-18L in the vanilla Post-LN. With the B2T connection, its training succeeded and it outperformed Pre-LN in the 18L-18L configuration. Figure 6 shows the negative log-likelihood (NLL) values of all methods when we regard newstest2013 as validation data. This figure indicates that the NLLs of Pre-LN are worse than those of the other methods. These results demonstrate that our modification enabled the stacking of many layers
+
+<table><tr><td>Method</td><td>2010</td><td>2011</td><td>2012</td><td>2013</td><td>2014</td><td>2015</td><td>2016</td><td>Average</td></tr><tr><td colspan="9">Enc-Dec: 100L-100L</td></tr><tr><td>Post-LN</td><td colspan="7">Training failed</td><td>N/A</td></tr><tr><td>Pre-LN</td><td>24.81</td><td>22.67</td><td>23.15</td><td>26.98</td><td>28.42</td><td>30.50</td><td>34.53</td><td>27.29</td></tr><tr><td>B2T connection</td><td>25.26</td><td>23.27</td><td>23.72</td><td>27.50</td><td>29.33</td><td>31.57</td><td>35.37</td><td>28.00</td></tr></table>
+
+without harm to its performance such as Pre-LN.
+
+In the comparison with the recent methods, B2T connection outperformed them with respect to the averaged BLEU score. This result implies that our modification is superior to the recent methods. To make our findings more reliable, we also conduct a comparison with the recent methods on the summarization task.
+
+Table 2 shows results in a much deeper configuration: 100L-100L. This table also indicates that B2T connection stabilized the training and outperformed Pre-LN. Appendix C describes the details of this 100L-100L configuration and shows a comparison with the latest method, DeepNet (Wang et al., 2022).
+
+# 6.2 Abstractive Summarization
+
+# 6.2.1 Dataset
+
+The abstractive summarization task is one of the most famous sequence-to-sequence problems in NLP. In this study, we conduct the experiment on the headline generation task, which is the task of generating a headline from a given sentence (Rush et al., 2015). We used headline-sentence pairs extracted from Annotated English Gigaword (Napoles et al., 2012) by Rush et al. (2015). This dataset contains 3.8M headline-sentence pairs as the training set and 1951 pairs as the test set. In addition, we used 13M additional headline-sentence pairs extracted from REALNEWS (Zellers et al., 2019) and NewsCrawl (Barrault et al., 2019) for training deep Transformers, following Takase and Kiyono (2021). We applied BPE (Sennrich et al., 2016) to construct a vocabulary set. As in the machine translation experiments, we set the number of BPE merge operations to 32K and shared the vocabulary between the encoder and decoder sides.
+
+# 6.2.2 Methods
+
+We compare Post-LN, Pre-LN, and B2T connection Transformers in the same manner as in Section 6.1. In addition, we compare DLCL, Admin, T-Fixup, RealFormer, and DeepNet because it
+
+Table 2: BLEU scores on WMT newestest2010-2016 and their averages in the 100L-100L configuration.  
+
+<table><tr><td>Method</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td colspan="4">Enc-Dec: 6L-6L</td></tr><tr><td>Post-LN</td><td>38.57</td><td>19.37</td><td>35.79</td></tr><tr><td>Pre-LN</td><td>38.27</td><td>19.29</td><td>35.39</td></tr><tr><td>DLCL (Wang et al., 2019)</td><td>38.13</td><td>18.49</td><td>35.00</td></tr><tr><td>Admin (Liu et al., 2020)</td><td>37.96</td><td>18.93</td><td>35.05</td></tr><tr><td>T-Fixup (Huang et al., 2020)</td><td>38.11</td><td>19.13</td><td>35.32</td></tr><tr><td>RealFormer (He et al., 2021)</td><td>38.30</td><td>19.32</td><td>35.46</td></tr><tr><td>DeepNet (Wang et al., 2022)</td><td>38.27</td><td>18.89</td><td>35.34</td></tr><tr><td>B2T connection</td><td>38.43</td><td>19.37</td><td>35.72</td></tr><tr><td colspan="4">Enc-Dec: 18L-18L</td></tr><tr><td>Post-LN</td><td colspan="3">Training failed</td></tr><tr><td>Pre-LN</td><td>38.97</td><td>19.94</td><td>35.99</td></tr><tr><td>DLCL (Wang et al., 2019)</td><td>38.25</td><td>19.44</td><td>35.57</td></tr><tr><td>Admin (Liu et al., 2020)</td><td>39.10</td><td>20.08</td><td>36.30</td></tr><tr><td>T-Fixup (Huang et al., 2020)</td><td>39.15</td><td>19.97</td><td>36.34</td></tr><tr><td>RealFormer (He et al., 2021)</td><td>39.22</td><td>20.12</td><td>36.49</td></tr><tr><td>DeepNet (Wang et al., 2022)</td><td>39.27</td><td>19.97</td><td>36.41</td></tr><tr><td>B2T connection</td><td>39.61</td><td>20.28</td><td>36.66</td></tr></table>
+
+Table 3: F1 based ROUGE-1, 2, and L scores (columns headed R-1, R-2, and R-L, respectively) on headline generation (Rush et al., 2015).
+
+would be premature to conclude that our modification is more effective than those methods from the results of experiments on the machine translation task alone. We set the numbers of layers of encoders and decoders to 6L-6L and 18L-18L as the base and deep configurations, respectively.
+
+# 6.2.3 Results
+
+Table 3 shows the ROUGE-1, 2, and L scores achieved by each method on the test set. Since these scores are computed by n-gram overlapping between the generated and correct headlines, a higher score represents better performance.
+
+In the 6L-6L configuration, Post-LN achieved better performance than Pre-LN. Thus, Post-LN outperformed Pre-LN on the headline generation task if training succeeded. Moreover, B2T connection achieved scores comparable to those of Post-LN.
+
+In the 18L-18L configuration, the training of Post-LN failed. In contrast, the training of B2T connection succeeded, and this method outperformed Pre-LN. Thus, our modification is more suitable than Pre-LN for training deep Transformers to perform the headline generation task.
+
+B2T connection outperformed the recent methods in the 6L-6L configuration and achieved the best ROUGE scores in the 18L-18L configuration. According to the results on both the machine translation and headline generation tasks, B2T connection achieved performance that was better than, or comparable to, that of previous methods. It is worth emphasizing that, in addition to the performance, our modification does not incur additional computational costs, such as those incurred by DLCL and Admin.
+
+# 6.3 Language Model
+
+In addition to encoder-decoders, we investigate the effect of our B2T connection when used in the decoder side only, i.e., a neural language model. Because recent pre-trained models, such as the GPT series, are language models trained on a large amount of training data, experimental results in this section give an insight for pre-trained models.
+
+# 6.3.1 Dataset
+
+We used WikiText-103 (Merit et al., 2017), which consists of a large number of tokens. The training, validation, and test sets contain 103M, 0.2M, and 0.2M tokens, respectively. The vocabulary set contains 0.3M words.
+
+# 6.3.2 Methods
+
+We used a Transformer with adaptive input representations (Baevski and Auli, 2019), which is implemented in fair seq, as the base architecture in this experiment. For the base configuration, we stacked 6 layers, in the same manner as in the machine translation and summarization experiments. For the deep configuration, we used 16 layers, following Baevski and Auli (2019). For the dimensions of internal layers, we used the same values as those used by Baevski and Auli (2019). We compare Post-LN, Pre-LN, and B2T connection.
+
+# 6.3.3 Results
+
+Table 4 shows perplexities of each method on the validation and test sets of WikiText-103. Since the perplexity is computed based on the negative log-likelihood, a smaller value corresponds to better performance. The upper part of this table indicates that, with 6 layers, Post-LN and our B2T connection outperformed Pre-LN. When we stacked 16 layers, the training of Post-LN failed, but B2T connection achieved better performance than Pre-LN. These results are consistent with results on
+
+<table><tr><td>Method</td><td>Valid</td><td>Test</td></tr><tr><td colspan="3">Dec: 6L</td></tr><tr><td>Post-LN</td><td>20.24</td><td>21.22</td></tr><tr><td>Pre-LN</td><td>20.98</td><td>21.93</td></tr><tr><td>B2T connection</td><td>20.50</td><td>21.47</td></tr><tr><td colspan="3">Dec: 16L</td></tr><tr><td>Post-LN</td><td colspan="2">Training failed</td></tr><tr><td>Pre-LN</td><td>18.53</td><td>19.24</td></tr><tr><td>B2T connection</td><td>18.38</td><td>19.20</td></tr></table>
+
+Table 4: Perplexities on WikiText-103 (Merit et al., 2017).  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Dev</td><td colspan="2">Test</td></tr><tr><td>Clean</td><td>Other</td><td>Clean</td><td>Other</td></tr><tr><td colspan="5">Enc-Dec: 6L-6L</td></tr><tr><td>Post-LN</td><td>3.78</td><td>8.76</td><td>4.19</td><td>8.74</td></tr><tr><td>Pre-LN</td><td>3.89</td><td>9.69</td><td>4.22</td><td>9.65</td></tr><tr><td>B2T connection</td><td>3.69</td><td>8.97</td><td>3.86</td><td>8.94</td></tr><tr><td colspan="5">Enc-Dec: 12L-6L</td></tr><tr><td>Post-LN</td><td></td><td colspan="3">Training failed</td></tr><tr><td>Pre-LN</td><td>3.21</td><td>7.91</td><td>3.49</td><td>8.22</td></tr><tr><td>B2T connection</td><td>3.26</td><td>7.74</td><td>3.48</td><td>7.68</td></tr></table>
+
+Table 5: Word error rates of each method on LibriSpeech.
+
+the machine translation and summarization tasks. Thus, our modification enables the training of deep Transformers for language modeling, and it is more effective than Transformers with Pre-LN.
+
+# 6.4 Automatic Speech Recognition
+
+In addition to experiments on natural language processing tasks, we conduct an experiment on another modality, ASR.
+
+# 6.4.1 Dataset
+
+We used LibriSpeech (Panayotov et al., 2015), which is the standard English ASR benchmark dataset. The dataset contains 1,000 hours of English speech extracted from audiobooks. We used the standard splits of LibriSpeech: we used all available training data for training and two configurations ('clean' and 'other') of development sets and test sets for evaluation. We applied the same pre-processing as that used by Wang et al. (2020). We constructed a vocabulary set for the decoder-side with SentencePiece (Kudo and Richardson, 2018) by setting the vocabulary size to 10,000. To obtain speech features, we used torchaudio<sup>9</sup>.
+
+# 6.4.2 Methods
+
+We used the Transformer-based speech-to-text model described in Wang et al. (2020) as the base
+
+architecture in this experiment. This model contains a convolutional layer to construct an embedding for the encoder-side but the other parts are identical to the Transformers used on the machine translation and summarization tasks. We used the same dimensions as those of T-Md, described in Wang et al. (2020). We set the numbers of layers to 6L-6L and 12L-6L as the base and deep configurations, respectively, because Wang et al. (2020) stacked many layers on the encoder-side only. We compare Post-LN, Pre-LN, and B2T connection.
+
+# 6.4.3 Results
+
+Table 5 shows the word error rates (WERs) of each method on each set. A smaller value of WER corresponds to better performance. The upper part of this table indicates that Post-LN and B2T connection outperformed Pre-LN on all sets in the 6L-6L configuration. The lower part of the table shows that B2T connection succeeded in training and achieved performance that was better than (or comparable to) that of Pre-LN in the 12L-6L configuration<sup>10</sup>. These results are consistent with those of the other experiments in this study.
+
+# 7 Related Work
+
+Layer normalization (Ba et al., 2016) is a useful technique for training neural networks but its mechanism has been unclear (Xu et al., 2019). The Transformer, which is the standard architecture for various tasks, also contains layer normalizations. The original Transformer architecture adopted the Post-LN configuration (Vaswani et al., 2017). However, recent Transformer implementations have adopted Pre-LN configurations (Klein et al., 2017; Vaswani et al., 2018; Ott et al., 2019; Baevski and Auli, 2019).
+
+To construct deep Transformers that achieve better performance, recent studies have focused on the behavior of layer normalizations. Wang et al. (2019) indicated the difficulty of training deep Transformers with Post-LN due to the vanishing gradient problem, and demonstrated that Pre-LN enables the stacking of many layers through machine translation experiments. In addition, they proposed a method to connect all layers to increase the effectiveness of deep Transformers. Bapna
+
+et al. (2018) and Dou et al. (2018) also proposed such connection methods to stack many layers. He et al. (2021) introduced additional connections into attention sub-layers to improve the performance. Xiong et al. (2020) explored the relation between the warm-up strategy and layer normalizations in Transformers. Through theoretical and empirical analyses, they indicated that Post-LN requires the warm-up strategy to stabilize the training.
+
+Liu et al. (2020) analyzed the training dynamics of Post-LN and Pre-LN Transformers. They then proposed Admin, which consists of additional weight parameters to control the variances of outputs from each sub-layer. In contrast, we indicated that we can stabilize the training of Post-LN Transformers by adding only a residual connection that skips over layer normalizations that cause the vanishing gradient.
+
+Some studies have proposed initialization methods to make the training of deep neural networks stable (Zhang et al., 2019a,b; Huang et al., 2020). Zhang et al. (2019a) proposed the depth-scaled initialization to prevent the vanishing gradient problem in Transformers. Zhang et al. (2019b) proposed the fixed-update initialization to remove normalizations in neural networks. Inspired by these studies, Huang et al. (2020) proposed T-Fixup, which enables both warm-up and layer normalizations to be removed from Transformers. In addition to the initialization scheme, Wang et al. (2022) introduced weights into residual connections before layer normalizations, following Liu et al. (2020).
+
+# 8 Conclusion
+
+In this study, we addressed the stability of training Post-LN Transformers. Through theoretical and empirical analyses, we indicated that layer normalizations cause the unstable training when many layers are stacked. In addition, we investigated the reason for the different performance of Pre-LN and Post-LN by transformations of each layer. We introduced B2T connection to prevent the vanishing gradient while preserving the advantage of Post-LN. We conducted experiments on various tasks. The experimental results led to the following three findings; 1, Post-LN achieved better performance than Pre-LN if its training succeeded. 2, Our modification enabled the training of deep Transformers (e.g., those with ten or more layers). 3, Our modification preserved the benefit of Post-LN, and therefore outperformed Pre-LN.
+
+# Limitations
+
+In this paper, we indicated that the vanishing gradient problem, caused by layer normalizations, makes the training of deep Post-LN Transformers unstable. We proposed the B2T connection to mitigate this vanishing gradient problem. However, the proposed B2T connection does not perfectly prevent the vanishing gradient, as shown in Figure 3. Therefore, the vanishing gradient might harm the training in extremely deep Transformers even if our B2T connection is used.
+
+In addition, this study depends on empirical observations. In particular, we provided little theoretical justification of the reason for Post-LN outperforming Pre-LN when training succeeds. However, as discussed in Appendix C, the method with a theoretical justification often collapses in some situations. Because the behavior of deep Transformers in various situations is not fully understood, we believe that it is important to provide empirical findings for our research field to progress.
+
+Although Appendix C includes a comparison between our B2T connection and the latest method, DeepNet (Wang et al., 2022), we could not investigate the behavior of all methods in the 100L-100L configuration because of our limited computational budgets. However, we are confident that we conducted sufficient experiments to verify our contributions.
+
+# Ethics Statement
+
+The proposed method helps to construct deep Transformers. As discussed in Strubell et al. (2019) and Schwartz et al. (2019), such deep neural networks consume substantial amounts of energy. In fact, as discussed in Appendix A.2, we spent a large amount of computational resources on our experiments. Therefore, we also need to explore methods of improving energy efficiency while maintaining the good performance achieved by stacking many layers.
+
+With respect to ethical considerations, the datasets used in our experiments are publicly available. LibriSpeech (Panayotov et al., 2015) is derived from audiobooks. The other datasets are mainly constructed from newswire texts and Wikipedia. Thus, in our understanding, our used datasets do not contain any personally identifiable information or offensive contents.
+
+# Acknowledgements
+
+We thank the anonymous reviewers for their useful suggestions. A part of this work was supported by JSPS KAKENHI Grant Number JP21K17800 and JST ACT-X Grant Number JPMJAX200I. The work of Jun Suzuki was partly supported by JST Moonshot R&D Grant Number JPMJMS2011 (fundamental research). We thank Edanz for editing a draft of this manuscript.
+
+# References
+
+Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer normalization.  
+Alexei Baevski and Michael Auli. 2019. Adaptive input representations for neural language modeling. In Proceedings of the 7th International Conference on Learning Representations (ICLR).  
+Ankur Bapna, Mia Chen, Orhan First, Yuan Cao, and Yonghui Wu. 2018. Training deeper neural machine translation models with transparent attention. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3028-3033.  
+Loic Barrault, Ondrej Bojar, Marta R. Costa-jussa, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, Christof Monz, Mathias Muller, Santanu Pal, Matt Post, and Marcos Zampieri. 2019. Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation (WMT), pages 1-61.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), pages 1877-1901.  
+Zi-Yi Dou, Zhaopeng Tu, Xing Wang, Shuming Shi, and Tong Zhang. 2018. Exploiting deep representations for neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4253-4262.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016a. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778.
+
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016b. Identity mappings in deep residual networks. In 14th European Conference on Computer Vision, pages 630-645.  
+Ruining He, Anirudh Ravula, Bhargav Kanagal, and Joshua Ainslie. 2021. RealFormer: Transformer likes residual attention. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 929-943.  
+Xiao Shi Huang, Felipe Perez, Jimmy Ba, and Maksims Volkovs. 2020. Improving transformer optimization through better initialization. In Proceedings of the 37th International Conference on Machine Learning, volume 119, pages 4475-4483.  
+Sergey Ioffe and Christian Szegedy. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning (ICML), volume 37, pages 448-456.  
+Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander Rush. 2017. OpenNMT: Open-source toolkit for neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL), pages 67-72.  
+Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 66-71.  
+Liyuan Liu, Xiaodong Liu, Jianfeng Gao, Weizhu Chen, and Jiawei Han. 2020. Understanding the difficulty of training transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5747-5763.  
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. Pointer Sentinel Mixture Models. In Proceedings of the 5th International Conference on Learning Representations (ICLR).  
+Courtney Naples, Matthew Gormley, and Benjamin Van Durme. 2012. Annotated Gigaword. In Proceedings of the Joint Workshop on Automatic Knowledge Base Construction and Web-scale Knowledge Extraction (AKBC-WEKEX), pages 95-100.  
+Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)*, pages 48-53.  
+Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. 2018. Scaling neural machine translation. In Proceedings of the Third Conference on Machine Translation (WMT), pages 1-9.
+
+Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206-5210.  
+Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation (WMT), pages 186-191.  
+Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A Neural Attention Model for Abstractive Sentence Summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 379-389.  
+Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. 2019. Green AI. CoRR, abs/1907.10597.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL), pages 1715-1725.  
+Rupesh K Srivastava, Klaus Greff, and Jürgen Schmidhuber. 2015. Training very deep networks. In Advances in Neural Information Processing Systems 28 (NIPS), pages 2377-2385.  
+Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 3645-3650.  
+Sho Takase and Shun Kiyono. 2021. Rethinking perturbations in encoder-decoders for fast training. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 5767-5780.  
+Sho Takase and Naoaki Okazaki. 2019. Positional encoding to control output sequence length. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), pages 3999-4004.  
+Ashish Vaswani, Samy Bengio, Eugene Brevdo, Francois Chollet, Aidan Gomez, Stephan Gouws, Llion Jones, Lukasz Kaiser, Nal Kalchbrenner, Niki Parmar, Ryan Sepassi, Noam Shazeer, and Jakob Uszkoreit. 2018. Tensor2Tensor for neural machine translation. In Proceedings of the 13th Conference of the Association for Machine Translation in the Americas, pages 193-199.  
+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 (NIPS), pages 5998-6008.
+
+Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Pino. 2020. Fairseq S2T: Fast speech-to-text modeling with fairseq. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing (AACL-IJCNLP), pages 33-39.  
+Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Dongdong Zhang, and Furu Wei. 2022. Deepnet: Scaling transformers to 1,000 layers.  
+Qiang Wang, Bei Li, Tong Xiao, Jingbo Zhu, Changliang Li, Derek F. Wong, and Lidia S. Chao. 2019. Learning deep transformer models for machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 1810-1822.  
+Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tie-Yan Liu. 2020. On layer normalization in the transformer architecture. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 10524-10533.  
+Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. 2019. Understanding and improving layer normalization. In Advances in Neural Information Processing Systems (NeurIPS), volume 32.  
+Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. In Advances in Neural Information Processing Systems 32 (NeurIPS), pages 9054-9065.  
+Biao Zhang, Ivan Titov, and Rico Sennrich. 2019a. Improving deep transformer with depth-scaled initialization and merged attention. 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 898-909.  
+Hongyi Zhang, Yann N. Dauphin, and Tengyu Ma. 2019b. Fixup initialization: Residual learning without normalization. In Proceedings of the 7th International Conference on Learning Representations (ICLR).
+
+# A Details of Experimental Settings
+
+# A.1 Hyper-parameters
+
+As described in Section 6, our hyper-parameters follow those used in previous studies. Table 6 shows hyper-parameters used for each experiment. For fair comparisons, we used the same hyperparameters for all methods except T-Fixup. For T-Fixup, we used hyper-parameters reported in Huang et al. (2020) to prevent divergence.
+
+# A.2 Computational Resources
+
+We mainly used NVIDIA Tesla P100 GPUs for most of our experiments. Table 7 shows the number of GPUs and the computational time used to construct one model in our experiments. For the 100L-100L configuration, described in Section 6.1, we used 24 Tesla V100 GPUs and spent approximately 120 hours to train one model.
+
+# B Supplementary of Gradient Norms of Each Location
+
+For gradient norms of each part in a layer, we check 1st and 9th decoders in addition to the 18th decoder for the 18L-18L Post-LN Transformer encoder-decoder as shown in Figure 4. Figure 7 shows the gradient norms of each part. This figure shows that the gradient norms decrease drastically through layer normalizations in the same manner as they do in the 18th decoder (Figure 4). Therefore, the vanishing gradient problem in Post-LN Transformers is probably caused by layer normalizations.
+
+# C Details of the 100L-100L Configuration
+
+# C.1 Regularizations during the Training
+
+As reported in Section 1, we constructed 100L-100L Transformers with widely-used WMT English-to-German dataset. In the preliminary experiments, we found that regularization is the key to preventing overfitting and achieving high performance in this situation. Figure 8 shows the NLL values of Pre-LN and B2T connection on validation data in the 36L-36L configuration when we used the same hyper-parameters as those used in 6L-6L and 18L-18L configurations. As this figure shows, the NLL values began to increase from the middle of training, and thus the overfitting occurred. In addition, the use of the same hyper-parameters as 6L-6L and 18L-18L makes it difficult to improve the performance of deeper configurations. Figure 9 shows the best NLL values on validation data when
+
+we varied the number of layers: 6L-6L, 12L-12L, 18L-18L, 36L-36L, and  $50\mathrm{L} - 50\mathrm{L}^{11}$ . This figure indicates that adding more layers to the 18L-18L configuration did not improve the performance.
+
+To prevent overfitting during the training of 100L-100L Transformers, we increased the dropout rate from 0.3 to 0.5. In addition, we used word dropout, as described in Takase and Kiyono (2021). We set the word dropout rate to 0.1 for the encoder and decoder. We multiplied the initial parameter values, except those for embeddings, by 0.1. We set the gradient clipping to 0.1. Finally, we decreased the number of updates from 50K to 25K. These regularization techniques prevented overfitting and achieved better performance than 18L-18L, as described in Section 6.1.
+
+# C.2 Comparison with DeepNet
+
+As described in Section 7, various studies have attempted to stabilize the training of deep Transformers. Each study indicated the effectiveness of their proposed method empirically, and some have provided theoretical justifications. However, Wang et al. (2022) demonstrated that the training of previous methods except DeepNet failed in a much deeper configuration than normally used, i.e., 100L-100L. Then, can we conclude that DeepNet is a silver bullet for deep Transformers? It is difficult to reach this conclusion because the training of DeepNet also fails in some configurations. For example, when we train deep Transformers, we might decrease the batch size because the trainable parameters occupy most of the GPU memories. When we tried this, the NLL value of DeepNet on validation data diverged, as shown in Figure 10. In other words, the training of DeepNet failed. In contrast, the training of our B2T connection succeeded in this situation. This result implies that there are problems in the training of deep Transformers that have not been solved in previous studies. Therefore, we believe that we should continue to add the empirical findings about new techniques, including B2T connection, to those of previous studies.
+
+# D B2T Connection without Layer Normalization
+
+In addition to B2T connection, we also consider a further modification to prevent the vanishing gra
+
+<table><tr><td>Params</td><td>Machine Translation</td><td>Abstractive Summarization</td><td>Language Model</td><td>ASR</td></tr><tr><td>Hidden dim size</td><td>512</td><td>512</td><td>1024</td><td>512</td></tr><tr><td>FFN dim size</td><td>2048</td><td>2048</td><td>4096</td><td>2048</td></tr><tr><td>Attention heads</td><td>8</td><td>8</td><td>8</td><td>8</td></tr><tr><td>Learning rate</td><td>0.001</td><td>0.001</td><td>0.001</td><td>0.001</td></tr><tr><td>Scheduler</td><td>inverse sqrt</td><td>inverse sqrt</td><td>inverse sqrt</td><td>inverse sqrt</td></tr><tr><td>Adam β</td><td>(0.9, 0.98)</td><td>(0.9, 0.98)</td><td>(0.9, 0.98)</td><td>(0.9, 0.98)</td></tr><tr><td>Warmup updates</td><td>4K</td><td>4K</td><td>2K</td><td>4K</td></tr><tr><td>Max updates</td><td>50K</td><td>50K</td><td>50K</td><td>150K</td></tr><tr><td>Max tokens / GPU</td><td>3584</td><td>3584</td><td>1024</td><td>40K</td></tr></table>
+
+Table 6: Hyper-parameters used in our experiments.  
+
+<table><tr><td></td><td colspan="2">Machine Translation</td><td colspan="2">Abstractive Summarization</td><td colspan="2">Language Model</td><td colspan="2">ASR</td></tr><tr><td></td><td>6L-6L</td><td>18L-18L</td><td>6L-6L</td><td>18L-18L</td><td>6L</td><td>16L</td><td>6L-6L</td><td>12L-6L</td></tr><tr><td>#GPU</td><td>128</td><td>128</td><td>64</td><td>144</td><td>128</td><td>192</td><td>32</td><td>32</td></tr><tr><td>Time (hour)</td><td>5</td><td>13</td><td>4</td><td>17</td><td>4</td><td>7</td><td>22</td><td>34</td></tr></table>
+
+Table 7: The number of GPUs and computational time used to construct one model in our experiments.
+
+![](images/e6aa8afb8448c37bc093ae71567e9f6ebb5c319875060e35efda841937ba0b80.jpg)  
+(a) 1st decoder
+
+![](images/237364754199b31a7f01235f43aaeb4770481f077aecc49d064530b9b4360b09.jpg)  
+(b) 9th decoder
+
+![](images/5665a345c318451161d547321d683d85e922ef4e2cd2bbc252d134a794dcb91c.jpg)  
+Figure 7: Gradient norms of each part in the (a) 1st decoder and (b) 9th decoder of the 18L-18L Post-LN Transformer encoder-decoder on WMT English-to-German translation training data.
+
+![](images/936d275dc7f447d2e5f946bb734a1f1b976a415ea9ea21e386a89c045e4feac1.jpg)  
+Figure 8: Negative Log-Likelihood (NLL) values of PreLN and our proposed B2T connection on validation data (newstest2013) in 36L-36L. We used the same hyperparameters as those used in 6L-6L and 18L-18L.  
+Figure 9: The best Negative Log-Likelihood (NLL) values on validation data (newstest2013) when the total number of layers is varied. The total number of layers is divided equally between the encoder and decoder.
+
+dient problem. Because layer normalizations decrease gradients drastically, as described in Section
+
+3, removing layer normalizations may provide stable gradients during back-propagation. However,
+
+![](images/c6ec098e9b0bb89b04b3533e6c57d00a9ca4c746dd22a09f22a66b291c329302.jpg)  
+Figure 10: Negative Log-Likelihood (NLL) values of DeepNet and our proposed B2T connection on validation data (newstest2013) in 100L-100L with a small batch size.
+
+the values in the forward pass increase exponentially if layer normalizations are removed. Therefore, we introduce weights that prevent the explosive increase in the forward pass while mitigating the decreasing gradients in back-propagation, as an alternative to the layer normalization. To use this alternative, we replace Equation (5) with the following equation:
+
+$$
+\alpha x _ {i n p} + \beta \left(x _ {f f n} + \operatorname {F F N} \left(x _ {f f n}\right)\right). \tag {9}
+$$
+
+Through several experiments $^{12}$ , we found that the following values of  $\alpha$  and  $\beta$  are suitable:
+
+$$
+\alpha = \min  \left(\frac {N}{1 2}, N ^ {- 0. 1 5}\right), \tag {10}
+$$
+
+$$
+\beta = d ^ {- 0. 2}, \tag {11}
+$$
+
+where  $N$  is the number of layers and  $d$  is the dimension of the input vectors  $x_{inp}$ . For example,  $N$  is set to 12 and 6 in the encoder and decoder, respectively, in the 12L-6L configuration. Therefore, as the number of layers increases, the value of  $\alpha$  increases while  $N$  remains small (until  $N = 9$ ), and then  $\alpha$  starts to decrease. In short,  $\alpha$  prevents an explosive increase in the forward pass when we stack many layers.  $\beta$  decreases as the dimension  $d$  increases, and thus it prevents an explosive increase when a large dimension is used. By using Equation (9), we can remove all layer normalizations in internal layers. This solves the vanishing gradient problem caused by layer normalizations.
+
+Tables 8 and 9 shows the results of B2T connection without layer normalizations ("w/o LN") on
+
+the machine translation and summarization tasks. These results indicate that B2T connection without layer normalizations achieved scores comparable to those of B2T connection with layer normalizations. However, because the results of B2T connection without layer normalizations are slightly worse than those with layer normalizations, we recommend the use of B2T connection with layer normalizations.
+
+<table><tr><td>Method</td><td>2010</td><td>2011</td><td>2012</td><td>2013</td><td>2014</td><td>2015</td><td>2016</td><td>Average</td></tr><tr><td colspan="9">Enc-Dec: 6L-6L</td></tr><tr><td rowspan="2">B2T connection + w/o LN</td><td>24.12</td><td>21.93</td><td>22.29</td><td>26.31</td><td>26.84</td><td>29.48</td><td>34.73</td><td>26.53</td></tr><tr><td>24.17</td><td>22.07</td><td>22.24</td><td>25.83</td><td>26.96</td><td>29.70</td><td>34.42</td><td>26.48</td></tr><tr><td colspan="9">Enc-Dec: 18L-18L</td></tr><tr><td rowspan="2">B2T connection + w/o LN</td><td>24.75</td><td>22.88</td><td>23.09</td><td>27.12</td><td>28.82</td><td>30.99</td><td>33.64</td><td>27.33</td></tr><tr><td>24.47</td><td>22.37</td><td>22.58</td><td>27.04</td><td>28.34</td><td>30.49</td><td>34.38</td><td>27.10</td></tr></table>
+
+Table 8: BLEU scores of our modifications on WMT newstest2010-2016 and their averages.  
+
+<table><tr><td>Method</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td colspan="4">Enc-Dec: 6L-6L</td></tr><tr><td rowspan="2">B2T connection + w/o LN</td><td>38.43</td><td>19.37</td><td>35.72</td></tr><tr><td>38.63</td><td>19.75</td><td>35.77</td></tr><tr><td colspan="4">Enc-Dec: 18L-18L</td></tr><tr><td rowspan="2">B2T connection + w/o LN</td><td>39.61</td><td>20.28</td><td>36.66</td></tr><tr><td>39.29</td><td>20.01</td><td>36.48</td></tr></table>
+
+Table 9: F1 based ROUGE scores of our modifications on headline generation.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Limitations section, after Conclusion section
+
+A2. Did you discuss any potential risks of your work?
+
+Ethics Statement section, after Conclusion section
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Sections 1 and 8
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 5
+
+B1. Did you cite the creators of artifacts you used?
+
+Sections 6 and 7
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Ethics Statement section
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Not applicable. Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Ethics Statement section
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Section 6
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Section 6
+
+C Did you run computational experiments?
+
+Section 6 and Appendices A, B, C, and D
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Appendices A and C
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Not applicable. Left blank.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 6 and Appendices A
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank.

2023/B2T Connection_ Serving Stability and Performance in Deep Transformers/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3cf33a894692fc28133032acf2cc4f9cc869aae64c3d96ded1ceb5343fe36e0
+size 794492

2023/BLOCSUM_ Block Scope-based Source Code Summarization via Shared Block Representation/bdc012ed-0e12-4395-93e5-cef4ae1613b2_content_list.json

@@ -0,0 +1,2227 @@
+[
+    {
+        "type": "text",
+        "text": "BLOCSUM: Block Scope-based Source Code Summarization via Shared Block Representation",
+        "text_level": 1,
+        "bbox": [
+            124,
+            87,
+            872,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "YunSeok Choi, Hyojun Kim, Jee-Hyong Lee",
+        "bbox": [
+            309,
+            142,
+            692,
+            160
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "College of Computing and Informatics",
+        "bbox": [
+            341,
+            161,
+            658,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sungkyunkwan University",
+        "bbox": [
+            391,
+            177,
+            608,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Suwon, South Korea",
+        "bbox": [
+            415,
+            193,
+            586,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{ys.choi, rlagywns0213, john}@skku.edu",
+        "bbox": [
+            307,
+            210,
+            694,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Code summarization, which aims to automatically generate natural language descriptions of the source code, has become an essential task in software development for better program understanding. Abstract Syntax Tree (AST), which represents the syntax structure of the source code, is helpful when utilized together with the sequence of code tokens to improve the quality of code summaries. Recent works on code summarization attempted to capture the sequential and structural information of the source code, but they considered less the property that source code consists of multiple code blocks. In this paper, we propose BLOCSUM, BLOck scope-based source Code SUMmarization via shared block representation that utilizes block-scope information by representing various structures of the code block. We propose a shared block position embedding to effectively represent the structure of code blocks and merge both code and AST. Furthermore, we develop variant ASTs to learn rich information such as block and global dependencies of the source code. To prove our approach, we perform experiments on two real-world datasets, the Java dataset and the Python dataset. We demonstrate the effectiveness of BLOCSUM through various experiments, including ablation studies and a human evaluation.",
+        "bbox": [
+            144,
+            278,
+            460,
+            689
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            701,
+            258,
+            715
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A description of source code is very important in software development because it helps developers better understand programs. Advances in deep learning have enabled automatic code summarization and increased software maintenance efficiency. Previous approaches on automatic source code summarization can be categorized into sequence-based, structure-based, and hybrid approaches. Sequence-based approaches generated summaries by capturing the sequential information of source code (Iyer et al., 2016; Allamanis et al., 2016; Liang and Zhu, 2018; Hu et al., 2018b; Wei et al., 2019; Ye et al.,",
+        "bbox": [
+            112,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2020; Ahmad et al., 2020). They tokenized source code into a sequence of code tokens and encoded them using seq2seq models. Meanwhile, structure-based approaches used Abstract Syntax Tree (AST) to capture the structural information of code (Hu et al., 2018a; Fernandes et al., 2019; Shido et al., 2019; Harer et al., 2019; Zhang et al., 2019; LeClair et al., 2020; Liu et al., 2021; Lin et al., 2021; Allamanis et al., 2018; Wang and Li, 2021; Wu et al., 2021). They parsed the source code into the AST and utilized graph models such as Graph Neural Networks (GNNs). Some works flattened AST into a pre-order traversal sequence (Alon et al., 2019, 2018; LeClair et al., 2019; Wang and Li, 2021; Choi et al., 2021). Hybrid approaches utilized both the token sequence and the ASTs of codes (Wan et al., 2018; Wei et al., 2020; Zhang et al., 2020; Shi et al., 2021). They parallelly processed token sequences and ASTs with independent encoders and tried to merge them in the decoder.",
+        "bbox": [
+            507,
+            252,
+            884,
+            574
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, the existing approaches have some limitations. Sequence-based approaches treated the source code as a single statement, so they incorporated only the sequence information of the code without any structural information such as code blocks. Structural information is very important to understand code because a snippet of code can be considered as a hierarchy of blocks.",
+        "bbox": [
+            507,
+            577,
+            884,
+            706
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Structure-based approaches tried to catch such structural information of code, but they considered less sequential information of code. A snippet of code is a sequence, so sequential information is also important to understand code. Another problem is that they depended only ASTs to capture structural information of code. However, ASTs are not suitable for capturing structural information because they are syntax trees for grammatical purpose. Since the AST is a tree, there is only one path between every pair of nodes. Any two nodes in ASTs are connected, but with a relatively long path, which hinders capturing the structural rela",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11427",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 11427-11441",
+        "bbox": [
+            220,
+            945,
+            774,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            700,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2422250e06caa6d720fdf4d3b70133bba8410b8201061567976c51650d2c4f81.jpg",
+        "image_caption": [
+            "(a) Source code"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            86,
+            389,
+            268
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/b31eba5f051d5d0451bca247d96cd52f89d3395601994eb0d7d2007145119628.jpg",
+        "image_caption": [
+            "(b) Abstract Syntax Tree",
+            "Figure 1: An example of code snippet and its AST in the Java language. (a) The orange box denotes the code block and the number means the index of the code block. (b) The nodes in the orange rectangular box belong to the same block in the AST."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            405,
+            86,
+            873,
+            269
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tions of nodes. It causes difficulties in propagating structural information to distant nodes in the AST. Some structure-based approaches tried to alleviate these problems by providing additional graphs such as Control Flow Graph (CFG) and Program Dependence Graph (PDG), but the cost and time required to produce these graphs and integrate them into one graph are not negligible.",
+        "bbox": [
+            112,
+            367,
+            487,
+            495
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Hybrid approaches utilized both code and its AST. Since both sequential and structural information of code is necessary to understand code, the approaches showed higher performances than the previous ones. However, they failed to effectively merge two different types of information. They simply adopted independent encoders for each of them and tried to merge them in the decoder. Due to independent encoders, their representations are easy to be independent. It will make it hard to effectively merge both sequential information of code and structural information. Since the token sequence and the AST of code are just different descriptions of the same code, they need to be encoded so that they are correlated with each other.",
+        "bbox": [
+            112,
+            499,
+            487,
+            739
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these limitations mentioned above, we exploit the fact that the source code is a set of blocks consisting of multiple statements for a specific purpose. The code tokens in one code block are configured for the same purpose. As shown in Figure 1a, the code block if  $\\{\\dots\\}$  (orange) consists of statements that are executed when a certain condition is true. So, it needs to consider the information on which block each code token belongs to. To better capture structural information, we need to give each token not only positional",
+        "bbox": [
+            112,
+            741,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "information but also block positional information when encoding.",
+        "bbox": [
+            507,
+            367,
+            880,
+            398
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Since ASTs do not have enough information to capture the structural information of code, we need to modify ASTs. The additional information we try to add is block dependency and global dependency between nodes. For example, as shown in Figure 1b, the node \"max\" in the orange block of AST is only connected to its parent node, \"Statement\", but there exists implicit block dependency that the node \"max\" belongs to the same orange block as the node \"data\" in the purple dashed line. Furthermore, there exists implicit global dependency between nodes. For example, the node \"max\" in the orange box is the same variable as the nodes in the green dashed line. We need to add such information to ASTs.",
+        "bbox": [
+            507,
+            401,
+            882,
+            640
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Also, we need to match blocks in the token sequence with blocks in the AST. For example, code tokens (\"if\", \"({\", \"data\", ..., \"})\" (orange) in Figure 1a can be mapped to its corresponding nodes in the AST (\"If\", \"data\", ..., \"println\") (orange) in Figure 1b. Here, the block in the code and the block in the AST are the same part with an equal role, so we will utilize information on which block of code corresponds to which block of AST. Such information can make the code and AST correlated, and assist effective merging of two kinds of information.",
+        "bbox": [
+            505,
+            644,
+            882,
+            835
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose BLOCSUM, BLOck scope-based source Code SUMmarization via shared block representation that utilizes blockscope information of token sequences and ASTs. First, we propose the shared block position em",
+        "bbox": [
+            507,
+            839,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11428",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/abe0afa0053a5cf3dc04cdd783ca1d6e7c83350114a2b1e4ddb223c2118ebe8e.jpg",
+        "image_caption": [
+            "Figure 2: Overview of our proposed approach, BLOCSUM."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            92,
+            884,
+            374
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "beddings for effectively representing the structure of the code block and combining a correlation between the code and the AST encoders. Furthermore, we develop simple yet effective variants of ASTs to learn rich information such as block and global dependencies of the source code. To validate our approach, we perform experiments on the Java dataset and the Python dataset. We prove the superiority of BLOCSUM through various experiments including ablation studies and a human evaluation.",
+        "bbox": [
+            112,
+            428,
+            489,
+            587
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2 BLOCSUM",
+        "text_level": 1,
+        "bbox": [
+            112,
+            601,
+            252,
+            615
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present the details of our model. Figure 2 shows the overall architecture of BLOCSUM. We first introduce the shared block position embedding and the abstract syntax tree variants and explain the architecture of BLOCSUM in detail.",
+        "bbox": [
+            112,
+            627,
+            489,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.1 Shared Block Position Embedding",
+        "text_level": 1,
+        "bbox": [
+            112,
+            720,
+            430,
+            736
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We suppose that there are code tokens  $c_{i}$  in a code snippet  $\\mathbf{C} = \\{c_1, c_2, \\ldots\\}$  and AST nodes  $n_{i}$  in its AST sequence  $\\mathbf{N} = \\{n_1, n_2, \\ldots\\}$ . We aim to predict a summary given the code tokens and the AST nodes.",
+        "bbox": [
+            112,
+            741,
+            487,
+            820
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Code blocks are the basic structural components of source code. Usually, code tokens in a block are gathered for a certain purpose, so the tokens need to be identified that they are in the same block. To distinguish which blocks are, we assign indexes to each block in the order of blocks in the code. Then",
+        "bbox": [
+            112,
+            822,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "each token has a block position with an index of the block it belongs to. If the code token is in nested blocks, we choose the innermost block index as the block position.",
+        "bbox": [
+            507,
+            428,
+            882,
+            491
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In order to utilize the block information of each token, we develop the block position embedding layer. Tokens in the same block have the same block position embedding. The code token embedding for the code encoder,  $E_{c}$ , is defined as follows:",
+        "bbox": [
+            507,
+            492,
+            882,
+            571
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {c} (t) = W _ {c} (t) + P _ {c} (t) + B _ {c} (t) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            583,
+            880,
+            601
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "for code token  $t$ . In the equation 1,  $W_{c}$ ,  $P_{c}$ , and  $B_{c}$  are the word, position, and block position embedding layers for the code encoder, respectively. Two position embeddings,  $P_{c}$  and  $B_{c}$ , are learnable positional encoding.",
+        "bbox": [
+            507,
+            612,
+            882,
+            692
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We also combine the AST nodes with block position embeddings to ensure that nodes in the same block have identical block information when node representations are learned by the AST encoder. As with the block position of the code tokens, each node is assigned a block position value. Since the block in the AST is a sub-tree structure in the AST, the node has an index of the sub-tree to which it belongs. Nodes in the same block have the same block positional embedding as code tokens. The AST node embedding for the AST encoder,  $E_{s}$ , is defined as follows:",
+        "bbox": [
+            507,
+            693,
+            884,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {s} (n) = W _ {s} (n) + P _ {s} (n) + B _ {s} (n) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            897,
+            880,
+            914
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11429",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/3976b24a222534e207b3671b0fe020cadc8f8e2dc90d1d4d7fd593b10db397b2.jpg",
+        "image_caption": [
+            "(a) AST-original"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            84,
+            225,
+            137
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/92b79d1e047e4112dcd20804f097201b9a35a88eb80f3651f69679d95502c3d3.jpg",
+        "image_caption": [
+            "(b) AST-block"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            252,
+            84,
+            344,
+            142
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/5d60ee75777c12d2ca60f0bc860850fad1f6c2196e07b3fbd9abb821e469f8ef.jpg",
+        "image_caption": [
+            "(c) AST-global",
+            "Figure 3: We introduce three different types of the AST: (a) AST-original referred as the original AST, (b) AST-block connected by edges between two nodes which are in the same code block, (c) AST-global connected by edges between all the nodes in the AST."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            83,
+            478,
+            148
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "for node  $n$ . In Equation 2,  $W_{s}$ ,  $P_{s}$ , and  $B_{s}$  are the word, position, and block position embedding layers for the AST encoder, respectively. The position of a node is defined as the position in the pre-order traversal sequence of the AST. Two position embeddings,  $P_{s}$  and  $B_{s}$ , are also learnable.",
+        "bbox": [
+            112,
+            273,
+            487,
+            368
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "There are two different types of inputs, code and AST, but they are just different descriptions for the same snippet. If two encoders for code and AST learn representations for code tokens and AST nodes separately, their representations will be easy to be independent and very hard to effectively combine both sequential and structural information.",
+        "bbox": [
+            112,
+            370,
+            487,
+            480
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to correlate the representations learned by two encoders, we allow the encoders to share the block position embedding layer. If the code token and the AST node belong to the same block, they will have the same block position embedding value. That is, we utilize additional information on which parts of the code correspond to which parts of the AST to generate better representations. If the block position embedding layers are shared, the embeddings for a token,  $t$ , and a node,  $n$  are as follows:",
+        "bbox": [
+            112,
+            483,
+            487,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {c} (t) = W _ {c} (t) + P _ {c} (t) + B (t) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            667,
+            485,
+            683
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {s} (n) = W _ {s} (n) + P _ {s} (n) + B (n) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            174,
+            687,
+            485,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $B$  is the shared block position embedding layer.",
+        "bbox": [
+            112,
+            711,
+            485,
+            741
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Shared block position embedding can effectively merge the information from two encoders. Also, it helps the code and the AST encoders to capture the structure of source code by providing block information.",
+        "bbox": [
+            112,
+            744,
+            487,
+            822
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.2 Abstract Syntax Tree Variants",
+        "text_level": 1,
+        "bbox": [
+            112,
+            834,
+            398,
+            848
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The original AST is a structure in which a node is connected only to its parent and children nodes. It contains local information, but it does not include the entire structure information of the code.",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Two nodes in the same block have implicit block dependency. There is also global dependency that two nodes have the same meaning even if their hop distance is very long. To utilize rich structural information such as block and global dependencies, we develop a simple yet effective method to reconstruct variants of AST. We define three variants of the AST: AST- original, AST-block, and AST-global.",
+        "bbox": [
+            507,
+            84,
+            882,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "AST-original is the original AST, which contains information on the local dependency between nodes, as shown in Figure 3a. The variants of AST are graphs of which nodes are the same as those of AST, but of which links are different.",
+        "bbox": [
+            507,
+            214,
+            882,
+            292
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "AST-block is the first variant of AST. We obtain it by removing the edges in AST-original, and adding new edges between the nodes belonging to the same block to represent the block structure information, as shown in Figure 3b. It represents information on the block dependency between nodes in the AST.",
+        "bbox": [
+            507,
+            294,
+            882,
+            390
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "AST-global is the second variant of AST. As shown in Figure 3c, we fully connect all the nodes in the AST. It represents the global and complete dependency between nodes in the AST. If the pre-order sequence of AST-global is learned by graph model, the node representations in the sequence represent context information of the AST.",
+        "bbox": [
+            507,
+            392,
+            882,
+            502
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each of the three AST variants represents local, block, and global dependencies between nodes in the AST. When they are learned organically by the AST encoder, node representations will contain rich structural information of the AST.",
+        "bbox": [
+            507,
+            504,
+            882,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.3 BLOCSUM Architecture",
+        "text_level": 1,
+        "bbox": [
+            507,
+            596,
+            752,
+            611
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Code Encoder Our code transformer encoder consists of 6 transformer layers (Vaswani et al., 2017). Each layer of the code transformer encoder is composed of two layers: multi-head self-attention (Vaswani et al., 2017) and feed-forward network. And residual connection (He et al., 2016) and layer normalization (Ba et al., 2016) are performed on each two sub-layers. The transformer encoder captures the sequential and block information of the code tokens to which the shared block position embedding is added.",
+        "bbox": [
+            507,
+            619,
+            882,
+            795
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "AST Encoder We use Graph Attention Networks (GATs) (Velickovic et al., 2018) for learning three different AST variants defined above: AST-original, AST-block, and AST-global. Our AST encoder consists of 6 multiple GAT encoder layers. Each layer of the AST multiple GAT encoder consists of three GATs for each variant AST. Each GAT captures",
+        "bbox": [
+            507,
+            806,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11430",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "local dependencies for AST-original, block dependencies for AST-block, and global dependencies for AST-global, respectively.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For the  $l$ -th layer of the AST encoder, the process is performed as follows:",
+        "bbox": [
+            112,
+            133,
+            485,
+            162
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {l n} ^ {l} = G A T _ {l n} \\left(A _ {l n}, h _ {n} ^ {l - 1}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            206,
+            172,
+            391,
+            189
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {b n} ^ {l} = G A T _ {b n} \\left(A _ {b n}, h _ {n} ^ {l - 1}\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            205,
+            193,
+            485,
+            211
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {g n} ^ {l} = G A T _ {g n} (A _ {g n}, h _ {n} ^ {l - 1})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            203,
+            215,
+            396,
+            234
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $GAT_{ln}, GAT_{bn}, GAT_{gn}$  and  $A_{ln}, A_{bn}, A_{gn}$  denote the GAT layers for three variant ASTs and the adjacency matrices in the AST-original, AST-block, and AST-global, respectively. Especially, the GAT for AST-global is the same as self-attention for learning the context of all nodes in the AST.",
+        "bbox": [
+            112,
+            240,
+            485,
+            336
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, the three representations are combined and performed from residual connection and layer normalization by the following equation:",
+        "bbox": [
+            112,
+            337,
+            485,
+            385
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {n} ^ {l} = L N \\left(h _ {n} ^ {l - 1} + F F N \\left(h _ {l n} ^ {l}, h _ {b n} ^ {l}, h _ {g n} ^ {l}\\right)\\right) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            139,
+            393,
+            485,
+            412
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $h_n^l$  is the concatenated node embedding in the  $l$ -th layer of AST GAT encoder,  $LN$  denotes layer normalization, and  $FFN$  is a feed-forward network.",
+        "bbox": [
+            112,
+            419,
+            485,
+            482
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "With the deep AST encoder layers, the node representations combine and propagate the local, block, and global information of the AST.",
+        "bbox": [
+            112,
+            483,
+            487,
+            532
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Summary Decoder The summary transformer decoder consists of 6 transformer decoder layers (Vaswani et al., 2017). Code token representations are learned with sequence and block information in the code transformer encoder and node representations are learned with local, block, and global dependencies of the code in the AST encoder. Given the code and node representations learned from each encoder, the summary transformer decoder learns to predict the summary of the original code token by fusion of the code and the AST representations. The multi-head self-attention in the decoder is performed sequentially on the code representations and node representations.",
+        "bbox": [
+            112,
+            541,
+            487,
+            765
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, when the summary transformer decoder predicts the  $t$ -th words, the copy mechanism (See et al., 2017) is applied to directly use the code tokens and AST nodes.",
+        "bbox": [
+            112,
+            766,
+            487,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3 Experiment Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            841,
+            317,
+            857
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.1 Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            866,
+            206,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets We evaluate using the benchmarks of the two real-world datasets, the Java dataset (Hu",
+        "bbox": [
+            112,
+            887,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "et al., 2018b) and the Python dataset (Wan et al., 2018). The experiment datasets are divided into 69,708/8,714/8,714 and 55,538/18,505/18,502 for train-valid/test, respectively. For extracting the AST of each dataset, we used a java parser javalang in the Java dataset and a python parser ast in the Python dataset used by Wan et al. (2018). Refer to Appendix A for the statistics of the datasets in detail.",
+        "bbox": [
+            507,
+            84,
+            882,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hyper-parameters We set the maximum length of code, AST, and summary to be 200, 200, and 50, respectively. For training the model, we use Adam optimizer (Kingma and Ba, 2015). We set the minibatch size as 80. The maximum training epoch is 100, and if the performance does not improve for 5 epochs, we stop early. Refer to Appendix C for the implementation details.",
+        "bbox": [
+            507,
+            239,
+            882,
+            367
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics We use BLEU (Papineni et al., 2002), METEOR (Banerjee and Lavie, 2005), and ROUGE-L (Lin, 2004) as metrics. We adopt S-BLEU, which indicates the average sentence-level BLEU score. Refer to Appendix B in detail.",
+        "bbox": [
+            507,
+            376,
+            882,
+            458
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines We adopt seq2seq models (Iyer et al., 2016; Hu et al., 2018a,b; Wei et al., 2019; Ahmad et al., 2020), graph2seq models (Eriguchi et al., 2016; Wan et al., 2018; Wu et al., 2021), hybrid models (Choi et al., 2021; Shi et al., 2021; Wu et al., 2021; Gong et al., 2022), and a pre-trained model CodeBERT (Feng et al., 2020) as the baselines. We fine-tuned the pre-trained model CodeBERT for code summarization.",
+        "bbox": [
+            507,
+            467,
+            882,
+            611
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2 Quantitative Result",
+        "text_level": 1,
+        "bbox": [
+            507,
+            624,
+            709,
+            639
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall Result Table 1 shows the overall performance of BLOCSUM and baselines on the Java and Python benchmark datasets. First, we can see that BLOCSUM improves the performance by 4.47 and 2.64 BLEU, 5.68 and 4.66 METEOR, and 4.63 and 3.85 ROUGE-L on the Java and Python datasets compared to the sequence-based approach, TransRel. In comparison with the structure-based approach, SITTransformer, the performance of BLOCSUM improves by 3.29 and 1.05 BLEU, 4.53 and 2.11 METEOR, and 3.84 and 2.23 ROUGE-L on the two datasets. The result shows that it is effective to capture the overall structure of code when AST is utilized together with the sequence of code tokens. Moreover, BLOCSUM performs better than hybrid approaches. Compared to GCNTransformer, the result shows that it",
+        "bbox": [
+            507,
+            645,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11431",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/eaee703ce4d7d3470df257816010dff289de95be1b1c3f14d5c8e835803872dc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">Java</td><td colspan=\"3\">Python</td></tr><tr><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td></tr><tr><td>CODE-NN (Iyer et al., 2016)</td><td>27.60</td><td>12.61</td><td>41.10</td><td>17.36</td><td>09.29</td><td>37.81</td></tr><tr><td>Tree2Seq (Eriguchi et al., 2016)</td><td>37.88</td><td>22.55</td><td>51.50</td><td>20.07</td><td>08.96</td><td>35.64</td></tr><tr><td>RL+Hybrid2Seq (Wan et al., 2018)</td><td>38.22</td><td>22.75</td><td>51.91</td><td>19.28</td><td>09.75</td><td>39.34</td></tr><tr><td>DeepCom (Hu et al., 2018a)</td><td>39.75</td><td>23.06</td><td>52.67</td><td>20.78</td><td>09.98</td><td>37.35</td></tr><tr><td>TL-CodeSum (Hu et al., 2018b)</td><td>41.31</td><td>23.73</td><td>52.25</td><td>15.36</td><td>08.57</td><td>33.65</td></tr><tr><td>Dual Model (Wei et al., 2019)</td><td>42.39</td><td>25.77</td><td>53.61</td><td>21.80</td><td>11.14</td><td>39.45</td></tr><tr><td>Transformer (Ahmad et al., 2020)</td><td>44.58</td><td>26.43</td><td>54.76</td><td>32.52</td><td>19.77</td><td>46.73</td></tr><tr><td>CodeBERT* (Feng et al., 2020)</td><td>41.32</td><td>27.42</td><td>55.33</td><td>30.72</td><td>21.53</td><td>49.93</td></tr><tr><td>GCNTransformer (Choi et al., 2021)</td><td>45.49</td><td>27.17</td><td>54.82</td><td>32.82</td><td>20.12</td><td>46.81</td></tr><tr><td>SiTTransformer (Wu et al., 2021)</td><td>45.76</td><td>27.58</td><td>55.58</td><td>34.11</td><td>21.11</td><td>48.35</td></tr><tr><td>CAST (Shi et al., 2021)</td><td>45.19</td><td>27.88</td><td>55.08</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SCRIPT (Gong et al., 2022)</td><td>46.89</td><td>28.48</td><td>56.69</td><td>34.00</td><td>20.84</td><td>48.15</td></tr><tr><td>BLOCSUM</td><td>49.05</td><td>32.11</td><td>59.42</td><td>35.16</td><td>23.22</td><td>50.58</td></tr></table>",
+        "bbox": [
+            137,
+            82,
+            858,
+            287
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "is more effective to capture both the sequential and structural information of the code considering the local, block, and global dependencies of AST rather than a flattened AST. Also, BLOCSUM considering the correlation between code and node representations performs better than Tripletopos using two independent encoders for the code and AST. Finally, we compared our approach with CodeBERT, a strong pre-trained program language model. BLOCSUM performs significantly better than CodeBERT trained on large code data. The result shows that our approach is more appropriate for code modeling than the pre-trained model in the code summarization task.",
+        "bbox": [
+            112,
+            351,
+            489,
+            575
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3 Qualitative Result",
+        "text_level": 1,
+        "bbox": [
+            112,
+            589,
+            302,
+            606
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Ablation Study we perform ablation studies to validate the effectiveness of shared block position embedding and AST variants on the Java and Python datasets.",
+        "bbox": [
+            112,
+            611,
+            487,
+            674
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "First, we design five models for comparison to verify the shared block position embedding: 1) not use block position embedding (unuse) 2) only use code block position embedding (code block emb) 3) only use AST block position embedding (ast block emb) 4) use separate block position embedding for code and AST (separate) 5) use shared block position embedding (share).",
+        "bbox": [
+            112,
+            677,
+            487,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In Table 2, code block emb and ast block emb have better performance than unuse. This result shows that block position embedding is effective in capturing the block information in each encoder. Also when both code and AST encoder use each separate block position embedding, we can see that it is more effective than only using one block po",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4194577a5effab64e8865e43061c769f68033c582d5efa2d5f81304202ef5233.jpg",
+        "table_caption": [
+            "Table 1: Comparison of our proposed model with the baseline models on the Java and Python datasets. We fine-tuned CodeBERT* with an input length of 200 and an output length of 50 for the two datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Block Pos Emb</td><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td></tr><tr><td colspan=\"4\">Java Dataset</td></tr><tr><td>unused</td><td>48.46</td><td>31.47</td><td>58.65</td></tr><tr><td>code block emb</td><td>48.53</td><td>31.56</td><td>58.63</td></tr><tr><td>ast block emb</td><td>48.54</td><td>31.77</td><td>58.92</td></tr><tr><td>separate</td><td>48.83</td><td>32.06</td><td>59.13</td></tr><tr><td>share</td><td>49.05</td><td>32.11</td><td>59.42</td></tr><tr><td colspan=\"4\">Python Dataset</td></tr><tr><td>unused</td><td>34.33</td><td>22.74</td><td>49.98</td></tr><tr><td>code block emb</td><td>34.50</td><td>22.84</td><td>50.10</td></tr><tr><td>ast block emb</td><td>34.56</td><td>22.83</td><td>50.08</td></tr><tr><td>separate</td><td>34.79</td><td>23.10</td><td>50.33</td></tr><tr><td>share</td><td>35.16</td><td>23.22</td><td>50.58</td></tr></table>",
+        "bbox": [
+            515,
+            348,
+            878,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Ablation study on shared block position embedding.",
+        "bbox": [
+            507,
+            548,
+            882,
+            577
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "sition embedding. Moreover, share has the best performance in comparison with other models. The shared block position embedding can learn the correlation between the code and AST encoders through the block-scope information rather than each separate block position embedding. Shared block position embedding not only effectively captures the structure of the code block, but also helps connect both the code and AST.",
+        "bbox": [
+            505,
+            608,
+            882,
+            751
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Second, we compare our approach with other combinations for verifying the effectiveness of variant ASTs: 1) AST-original (o) 2) AST-block (b), 3) AST-global (g).",
+        "bbox": [
+            505,
+            755,
+            882,
+            820
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Table 3, the results show that leveraging more structural information, such as block or global dependencies, performs better than modeling AST-original with only local dependency. Also, combining AST-original, AST-block, and AST-global has the best performance in comparison with",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11432",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/3d52032c664be3dd03269180d6db5c52b35895ae4a5283b865e7fda63f78a232.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Combination</td><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td></tr><tr><td colspan=\"4\">Java Dataset</td></tr><tr><td>o</td><td>48.39</td><td>31.73</td><td>58.98</td></tr><tr><td>g</td><td>48.41</td><td>31.63</td><td>58.40</td></tr><tr><td>o + g</td><td>48.76</td><td>31.87</td><td>59.00</td></tr><tr><td>o + b + g</td><td>49.05</td><td>32.11</td><td>59.42</td></tr><tr><td colspan=\"4\">Python Dataset</td></tr><tr><td>o</td><td>34.58</td><td>22.76</td><td>50.14</td></tr><tr><td>g</td><td>34.41</td><td>22.84</td><td>50.01</td></tr><tr><td>o + g</td><td>34.64</td><td>22.87</td><td>50.11</td></tr><tr><td>o + b + g</td><td>35.16</td><td>23.22</td><td>50.58</td></tr></table>",
+        "bbox": [
+            126,
+            80,
+            475,
+            249
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "combining other combinations. We demonstrate that utilizing AST variants helps learn rich information such as block and global dependencies of the source code.",
+        "bbox": [
+            112,
+            298,
+            489,
+            362
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Human Evaluation We performed a human evaluation on the Python dataset to demonstrate the quality of generated summaries. We randomly select 100 code snippets and ask three people with knowledge of the Python language to evaluate the summaries. They are CS graduate students with many years of experience in Python languages. Following the human evaluation metrics of (Choi et al., 2021), we ask them to evaluate the 3 following metrics: 1) Fluency (Quality of the summary & grammatically correct), 2) Relevance (Selection of the consistent content in source code), 3) Coverage (Selection of the important content in source code). We show pairs of summaries generated from BLOCSUM and the baseline fine-tuned (Feng et al., 2020) to the evaluators, and they select one of win, tie, and loss in three metrics for both results.",
+        "bbox": [
+            115,
+            370,
+            489,
+            643
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4 shows the results of human evaluation on the generated summaries on the Python dataset. The scores of fluency is lower, but the relevance and coverage are very higher than the baseline, CodeBERT. We analyzed the generated summaries of the two models and identified that BLOCSUM generates it similarly to the ground truth, reflecting the keyword of the code. CodeBERT, a pretrained language model, can generate more fluent and grammatical summaries, but the length is relatively short and a very plain summary with no keywords. The average tokens in the ground truth are 10.14, while the average tokens in summaries generated by BLOCSUM and CodeBERT are 9.91 and 8.16, respectively. We think that short sentences are more grammatically advantageous than long sentences. BLOCSUM has the highest tie in terms",
+        "bbox": [
+            112,
+            645,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b0ee04bce0d4002c443a7a2449423f755e6ebf72db4c707cd5be4287afbd8152.jpg",
+        "table_caption": [
+            "Table 3: Ablation study on combination of AST variants."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Fluency</td><td>Relevance</td><td>Coverage</td></tr><tr><td>Win</td><td>76</td><td>143</td><td>125</td></tr><tr><td>Tie</td><td>114</td><td>64</td><td>80</td></tr><tr><td>Loss</td><td>110</td><td>93</td><td>95</td></tr></table>",
+        "bbox": [
+            549,
+            80,
+            842,
+            146
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Human evaluation of the appropriateness of the generated adversarial example on the Python dataset.",
+        "bbox": [
+            507,
+            155,
+            880,
+            185
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "of fluency but the highest win in terms of relevance and coverage. The result means that BLOCSUM reflects more the core characteristic of the code than CodeBERT.",
+        "bbox": [
+            507,
+            216,
+            882,
+            279
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparison with the baselines Table 5 shows the summary examples generated from our proposed model on the Python dataset. The result on the Python dataset example shows BLOCSUM predicts the keywords \"wsgi\" and \"request\" by reflecting block-scope information. Although there is no dependency between two words in the code and the original AST, BLOCSUM utilizes three different types of AST variants and jointly learns structural dependency in three aspects to improve the performance of the model. Refer to Appendix D for more summary examples on the Java and Python datasets.",
+        "bbox": [
+            507,
+            300,
+            884,
+            508
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            532,
+            665,
+            546
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sequence-based Approaches Iyer et al. (2016) and Allamanis et al. (2016) proposed to use Long Short Term Memory (LSTM) and Convolutional Neural Networks (CNNs) for the source code summarization. Liang and Zhu (2018) proposed a tree-based recursive neural network to represent the syntax tree of code. Hu et al. (2018b) and Chen et al. (2021) summarized the source code with the APIs knowledge. Wei et al. (2019) used a dual training framework by training code summarization and code generation tasks. Also, Ye et al. (2020) considered the probabilistic correlation between the two tasks. Choi et al. (2020) proposed attention-based keyword memory networks for code summarization. Ahmad et al. (2020) proposed a Transformer model using a relative position. However, these approaches have limitations in that they did not explicitly incorporate the structural information of the source code, which is just as crucial as capturing the code semantics. Also, they did not learn the code block information because they learned the code as a sequence of tokens.",
+        "bbox": [
+            507,
+            565,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11433",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/c9de8851c7fdc3e9b273dea54dac3a1b81ebc719df53e860afaac0a2292c1f0a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Python Code</td><td>def simulate_request(app, method=&#x27;GET&#x27;, path=&#x27;/&#x27;, query_string=None, headers=None, body=None, fileRIAWRER=None, params=None, params_csv=True): if (not path.startswith(&#x27;/&#x27;)) : raise ValueError(&quot;path must start with &#x27;/&quot;) if (query_string and query_string.startswith(&#x27;?&#x27;)) : raise ValueError(&quot;query_string should not start with &#x27;?&#x27;) if (&#x27;?&#x27; in path): raise ValueError(&#x27;path may not contain a query string. Please use the query_string parameter instead.&#x27;) if (query_string is None): query_string = to_query_str(parms, comma delimitedLists=params_csv, prefix=False) env = helpers.create_environ(method=method, path=path, query_string=(query_string or ), headers=headers, body=body, file wraps=file wraps) srmock = StartResponseMock(   ) validator = wsgirefvalidate_validator(app) iterable = validator(env, smock) result = Result(iterable, smock.status, smockHeaders) return result</td></tr><tr><td>Ground Truth CodeBERT BLOCSUM</td><td>simulate a request to a wsgi application . simulate a wsgi environment . simulate a request to a wsgi request .</td></tr></table>",
+        "bbox": [
+            127,
+            82,
+            872,
+            349
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: A qualitative example on the Python dataset.",
+        "bbox": [
+            315,
+            357,
+            678,
+            373
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Structure-based Approaches Hu et al. (2018a) proposed an RNN-based model using the pre-order traversal sequence as input. Shido et al. (2019); Harer et al. (2019) adopted Tree-LSTM, TreeTransformer to encode tree-based inputs. LeClair et al. (2020) proposed encoded AST using graph neural networks and trained LSTM. Liu et al. (2021) proposed a retrieval augmented method with Graph Neural Network (GNN). Zhang et al. (2019) proposed AST-based Neural Network (ASTNN) for encoding the subtree. Lin et al. (2021) proposed Tree-LSTM to represent the split AST for code summarization. Li et al. (2021) leverage the retrieve-and-edit framework to improve the performance for code summarization. Allamanis et al. (2018), Wang and Li (2021), and Wu et al. (2021) tried to capture rich information using additional graphs such as CFG and PDG. But, these approaches considered only the structural information of the AST without considering the sequential information of the code token.",
+        "bbox": [
+            112,
+            398,
+            489,
+            734
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Hybrid Approaches Alon et al. (2019) leveraged the unique syntactic structure of programming languages by sampling paths in the AST of a code snippet. LeClair et al. (2019) proposed ast-attendgru model that combines code with structure from AST Also, Choi et al. (2021) proposed a model that combines Graph Convolution Network and Transformer using AST. Wu et al. (2021) incorporated a multiview graph matrix into the transformer model. Shi et al. (2021) tried to hierarchically split and",
+        "bbox": [
+            112,
+            758,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "reconstruct ASTs using Recursive Neural Network for learning the representation of the complete AST. Wan et al. (2018) used a deep reinforcement learning framework to consider an AST structure and code snippets. Gong et al. (2022) proposed a structural position method to augment the structural correlations between code tokens. But, they encoded two types of representations independently without correlation and did not consider merging them.",
+        "bbox": [
+            507,
+            398,
+            884,
+            543
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Zhang et al. (2020) proposed a retrieval-based approach using syntactic and semantic similarity for source code summarization. Liu et al. (2021) proposed a hybrid GNN using a retrieval augmented graph method. Wei et al. (2020) proposed a comment generation framework using AST, similar code, and exemplar from code. Choi et al. (2023) proposed a self-attention network that adaptively learns the structural and sequential information of code. But, they tried to model the code using more code information through retrieval methods.",
+        "bbox": [
+            507,
+            543,
+            884,
+            719
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            732,
+            640,
+            747
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we proposed BLOCSUM, BLOck scope-based source Code SUMmarization via shared block representation that utilizes blockscope information by representing various structures of the code block. We designed two methods using the fact that a code block is a fundamental structural component of the source code. We propose the first method, the shared block position embedding, for effectively representing the structure of the code block and merging a correlation",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11434",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "between the code and the AST encoders. Furthermore, we developed to reconstruct simple yet effective AST variants to learn rich information such as block and global dependencies of the source code. Experimental results demonstrated the effectiveness of BLOCSUM and confirmed the importance of block-scope information in the code.",
+        "bbox": [
+            112,
+            84,
+            490,
+            197
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            218,
+            220,
+            233
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we conducted an experiment on code summarization using two benchmark datasets, the Java dataset (Hu et al., 2018b) and the Python dataset (Wan et al., 2018). BLOCSUM may need to be tested for its generalizability to other program languages. We chose two program languages (Java and Python) that were easily parsed to map the block position of Code and AST. We believe that since other programming languages have similar syntactic structures, BLOCSUM should be able to achieve similar performance on them as well.",
+        "bbox": [
+            112,
+            250,
+            489,
+            428
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            448,
+            265,
+            462
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper proposes block scope-based source code summarization via shared block representation that utilizes block-scope information by representing various structures of the code block, which is beneficial to increase the efficiency of developers. The research conducted in this paper will not cause any ethical issues or have any negative social effects. The data used is all publicly accessible and is commonly used by researchers as a benchmark for program and language generation tasks. Our proposed method does not introduce any ethical or social bias or worsen any existing bias in the data.",
+        "bbox": [
+            112,
+            480,
+            489,
+            674
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            694,
+            285,
+            709
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No.2019-0-00421, AI Graduate School Support Program(Sungkyunkwan University)) This research was supported by the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience Program(IITP-2023-2020-0-01821) supervised by the IITP(Institute for Information & communications Technology Planning & Evaluation)",
+        "bbox": [
+            112,
+            726,
+            490,
+            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": [
+            "Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. 2020. A transformer-based approach for source code summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4998-5007, Online. Association for Computational Linguistics.",
+            "Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. 2018. Learning to represent programs with graphs. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.",
+            "Miltiadis Allamanis, Hao Peng, and Charles Sutton. 2016. A convolutional attention network for extreme summarization of source code. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 2091-2100. JMLR.org.",
+            "Uri Alon, Shaked Brody, Omer Levy, and Eran Yahav. 2019. code2seq: Generating sequences from structured representations of code. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.",
+            "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. 2018. code2vec: Learning distributed representations of code.",
+            "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normalization. *ArXiv preprint*, abs/1607.06450.",
+            "Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, Ann Arbor, Michigan. Association for Computational Linguistics.",
+            "Fuxiang Chen, Mijung Kim, and Jaegul Choo. 2021. Novel natural language summarization of program code via leveraging multiple input representations. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2510-2520, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "YunSeok Choi, JinYeong Bak, CheolWon Na, and JeeHyong Lee. 2021. Learning sequential and structural information for source code summarization. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 2842–2851, Online. Association for Computational Linguistics.",
+            "YunSeok Choi, Suah Kim, and Jee-Hyong Lee. 2020. Source code summarization using attention-based"
+        ],
+        "bbox": [
+            510,
+            107,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11435",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "keyword memory networks. In 2020 IEEE International Conference on Big Data and Smart Computing (BigComp), pages 564-570. IEEE.",
+            "YunSeok Choi, CheolWon Na, Hyojun Kim, and JeeHyong Lee. 2023. Readsum: Retrieval-augmented adaptive transformer for source code summarization. IEEE Access.",
+            "Akiko Eriguchi, Kazuma Hashimoto, and Yoshimasa Tsuruoka. 2016. Tree-to-sequence attentional neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 823-833, Berlin, Germany. Association for Computational Linguistics.",
+            "Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A pre-trained model for programming and natural languages. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1536-1547, Online. Association for Computational Linguistics.",
+            "Patrick Fernandes, Miltiadis Allamanis, and Marc Brockschmidt. 2019. Structured neural summarization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.",
+            "Zi Gong, Cuiyun Gao, Yasheng Wang, Wenchao Gu, Yun Peng, and Zenglin Xu. 2022. Source code summarization with structural relative position guided transformer. arXiv preprint arXiv:2202.06521.",
+            "Jacob Harer, Chris Reale, and Peter Chin. 2019. Tree-transformer: A transformer-based method for correction of tree-structured data. ArXiv preprint, abs/1908.00449.",
+            "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.",
+            "Xing Hu, Ge Li, Xin Xia, David Lo, and Zhi Jin. 2018a. Deep code comment generation. In 2018 IEEE/ACM 26th International Conference on Program Comprehension (ICPC), pages 200-20010. IEEE.",
+            "Xing Hu, Ge Li, Xin Xia, David Lo, Shuai Lu, and Zhi Jin. 2018b. Summarizing source code with transferred API knowledge. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden, pages 2269-2275. ijcai.org.",
+            "Srinivasan Iyer, Ioannis Konstas, Alvin Cheung, and Luke Zettlemoyer. 2016. Summarizing source code using a neural attention model. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2073-2083, Berlin, Germany. Association for Computational Linguistics.",
+            "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.",
+            "Alexander LeClair, Sakib Haque, Lingfei Wu, and Collin McMillan. 2020. Improved code summarization via a graph neural network. volume abs/2004.02843.",
+            "Alexander LeClair, Siyuan Jiang, and Collin McMillan. 2019. A neural model for generating natural language summaries of program subroutines. In 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE), pages 795-806. IEEE.",
+            "Jia Li, Yongmin Li, Ge Li, Xing Hu, Xin Xia, and Zhi Jin. 2021. Editsum: A retrieve-and-edit framework for source code summarization. In 2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE), pages 155-166. IEEE.",
+            "Yuding Liang and Kenny Qili Zhu. 2018. Automatic generation of text descriptive comments for code blocks. 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 5229-5236. AAAI Press.",
+            "Chen Lin, Zhichao Ouyang, Junqing Zhuang, Jianqiang Chen, Hui Li, and Rongxin Wu. 2021. Improving code summarization with block-wise abstract syntax tree splitting. In 2021 IEEE/ACM 29th International Conference on Program Comprehension (ICPC), pages 184-195. IEEE.",
+            "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.",
+            "Shangqing Liu, Yu Chen, Xiaofei Xie, Jing Kai Siow, and Yang Liu. 2021. Retrieval-augmented generation for code summarization via hybrid GNN. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.",
+            "Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11436",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Linguistics (Volume 1: Long Papers), pages 1073-1083, Vancouver, Canada. Association for Computational Linguistics.",
+            "Ensheng Shi, Yanlin Wang, Lun Du, Hongyu Zhang, Shi Han, Dongmei Zhang, and Hongbin Sun. 2021. CAST: Enhancing code summarization with hierarchical splitting and reconstruction of abstract syntax trees. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4053-4062, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Yusuke Shido, Yasuaki Kobayashi, Akihiro Yamamoto, Atsushi Miyamoto, and Tadayuki Matsumura. 2019. Automatic source code summarization with extended tree-lstm. volume abs/1906.08094.",
+            "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 Liò, and Yoshua Bengio. 2018. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.",
+            "Yao Wan, Zhou Zhao, Min Yang, Guandong Xu, Haochao Ying, Jian Wu, and Philip S Yu. 2018. Improving automatic source code summarization via deep reinforcement learning. In Proceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering, pages 397-407.",
+            "Yanlin Wang and Hui Li. 2021. Code completion by modeling flattened abstract syntax trees as graphs. In AAAI.",
+            "Bolin Wei, Ge Li, Xin Xia, Zhiyi Fu, and Zhi Jin. 2019. Code generation as a dual task of code summarization. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 6559-6569.",
+            "Bolin Wei, Yongmin Li, Ge Li, Xin Xia, and Zhi Jin. 2020. Retrieve and refine: exemplar-based neural comment generation. In 2020 35th IEEE/ACM International Conference on Automated Software Engineering (ASE), pages 349-360. IEEE.",
+            "Hongqiu Wu, Hai Zhao, and Min Zhang. 2021. Code summarization with structure-induced transformer. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 1078-1090, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Wei Ye, Rui Xie, Jinglei Zhang, Tianxiang Hu, Xiaoyin Wang, and Shikun Zhang. 2020. Leveraging code generation to improve code retrieval and summarization via dual learning. In WWW '20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, pages 2309-2319. ACM / IW3C2.",
+            "Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, and Xudong Liu. 2020. Retrieval-based neural source code summarization. In 2020 IEEE/ACM 42nd International Conference on Software Engineering (ICSE), pages 1385-1397. IEEE.",
+            "Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, Kaixuan Wang, and Xudong Liu. 2019. A novel neural source code representation based on abstract syntax tree. In 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE), pages 783-794. IEEE."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            326
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11437",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Statistics of Experiment Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            436,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For obtaining ASTs of the Java and Python dataset, we use the javalang<sup>1</sup> and ast<sup>2</sup> library, respectively. Also, we tokenize the source code and the AST to subtokens as the form CamelCase and snake-case.",
+        "bbox": [
+            112,
+            109,
+            489,
+            172
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/d71f66b9431d4561fad8989bfa23c73e6895962ef0b32a99c398df872bf42114.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Java</td><td>Python</td></tr><tr><td>Train</td><td>69,708</td><td>55,538</td></tr><tr><td>Valid</td><td>8,714</td><td>18,505</td></tr><tr><td>Test</td><td>8,714</td><td>18,502</td></tr><tr><td>Unique non-leaf nodes in ASTs</td><td>106</td><td>54</td></tr><tr><td>Unique leaf nodes in ASTs</td><td>57,372</td><td>101,229</td></tr><tr><td>Unique tokens in summaries</td><td>46,895</td><td>56,189</td></tr><tr><td>Avg. nodes in AST</td><td>131.72</td><td>104.11</td></tr><tr><td>Avg. tokens in summary</td><td>17.73</td><td>9.48</td></tr></table>",
+        "bbox": [
+            127,
+            181,
+            475,
+            319
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Statistics of Java dataset (Hu et al., 2018b) and Python dataset (Wan et al., 2018). For obtaining their corresponding ASTs, we use the javalang and ast libraries, respectively.",
+        "bbox": [
+            112,
+            328,
+            489,
+            387
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            414,
+            317,
+            430
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "BLEU(Papineni et al., 2002) is a Bilingual Evaluation Understudy to measure the quality of generated code summaries. The formula for computing BLEU is as follows:",
+        "bbox": [
+            112,
+            439,
+            489,
+            502
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nB L E U = B P \\cdot \\exp \\sum_ {n = 1} ^ {N} \\omega_ {n} \\log p _ {n}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            524,
+            411,
+            560
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "where  $p_n$  is the geometric average of the modified n-gram precisions,  $\\omega_{n}$  is uniform weights  $1 / N$  and BP is the brevity penalty.",
+        "bbox": [
+            112,
+            570,
+            487,
+            617
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "METEOR(Banerjee and Lavie, 2005) is used to measure how closely the metric scores match the human judgments about the quality of the translation. So unigram precision  $(P)$  and unigram recall  $(R)$  are computed and combined via a harmonic mean. The METEOR score is computed as follows:",
+        "bbox": [
+            112,
+            618,
+            489,
+            715
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM E T E O R = (1 - \\gamma \\cdot f r a g ^ {\\beta}) \\cdot \\frac {P \\cdot R}{\\alpha \\cdot P + (1 - \\alpha) \\cdot P}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            733,
+            472,
+            762
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "where  $\\text{frag}$  is the fragmentation fraction.  $\\alpha$ ,  $\\beta$ , and  $\\gamma$  are three penalty parameters whose default values are 0.9, 3.0, and 0.5, respectively.",
+        "bbox": [
+            112,
+            771,
+            487,
+            819
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "ROUGE-L(Lin, 2004) is used to apply Longest Common Subsequence in summarization evaluation. ROUGE-L used LCS-based F-measure to estimate the similarity between two summaries  $X$",
+        "bbox": [
+            112,
+            819,
+            489,
+            883
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "of length  $m$  and  $Y$  of length  $n$ , assuming  $X$  is a reference summary sentence, and  $Y$  is a candidate summary sentence, as follows:",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} R _ {l c s} = \\frac {L C S (X , Y)}{m}, P _ {l c s} = \\frac {L C S (X , Y)}{n} \\\\ F _ {l c s} = \\frac {\\left(1 + \\beta^ {2}\\right) R _ {l c s} P _ {l c s}}{R _ {l c s} + \\beta^ {2} P _ {l c s}} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            559,
+            151,
+            828,
+            208
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "where  $\\beta = P_{lcs} / R_{lcs}$  and  $F_{lcs}$  is the value of ROUGE-L.",
+        "bbox": [
+            507,
+            218,
+            882,
+            249
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Implementation Detail",
+        "text_level": 1,
+        "bbox": [
+            507,
+            262,
+            742,
+            279
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We conducted experiments on Ubuntu 18.04 with 4 2080 Ti GPUs. The environment of the sever supports python 3.9, Cuda 10.2, pytorch 1.9, and pytorch geometric 1.7.",
+        "bbox": [
+            507,
+            287,
+            882,
+            351
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The average training and inference time for BLOCSUM takes about 40 and 0.5 hours, respectively. BLOCSUM has about 76 million parameters.",
+        "bbox": [
+            507,
+            351,
+            882,
+            414
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/dd4666d556971846a08388928df89e4a93c2dc477da95bbea377893b9a742536.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyper-parameter</td><td>Size</td></tr><tr><td>the maximum length of code tokens</td><td>200</td></tr><tr><td>the maximum length of AST nodes</td><td>200</td></tr><tr><td>the maximum length of summary tokens</td><td>50</td></tr><tr><td>Embedding dimension</td><td>512</td></tr><tr><td>The number of Code Encoder layers</td><td>6</td></tr><tr><td>The number of AST Encoder layers</td><td>6</td></tr><tr><td>The number of Code Decoder layers</td><td>6</td></tr><tr><td>Head of Attention</td><td>8</td></tr><tr><td>batch size</td><td>80</td></tr><tr><td>train epoch</td><td>100</td></tr><tr><td>early stop</td><td>5</td></tr><tr><td>learning rate</td><td>0.0005</td></tr><tr><td>learning decay</td><td>0.99</td></tr><tr><td>beam size</td><td>5</td></tr></table>",
+        "bbox": [
+            531,
+            425,
+            860,
+            633
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Hyper-parameters of BLOCSUM.",
+        "bbox": [
+            549,
+            642,
+            840,
+            657
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://github.com/c2nes/javalang",
+        "bbox": [
+            134,
+            890,
+            349,
+            904
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://github.com/python/cpython/blob/master/Lib/ast.py",
+        "bbox": [
+            134,
+            904,
+            485,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "11438",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Examples of Java and Python Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            485,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "Java Code @ Override public String toString() { String result; result=super.toString(); if (m_CapabilitiesFilter != null) { initCapabilities(); if (m_Capabilities != null) { if (m_CapabilitiessupportsMaybe(m_CapabilitiesFilter) &&! m_Capabilitiessupports(m_CapabilitiesFilter)) { result=\"<html><font color  $\\equiv$  \"\" + MAYBE.'<sup>+</sup> + result + </font></i><html>\"; } else if(!m_Capabilitiessupports(m_CapabilitiesFilter)) { result=\"<html><font color  $\\equiv$  \"\" + NO.'<sup>+</sup> + result+ </font></i><html>\"; } } return result;   \n}   \nGround Truth return a string representation of this tree node .   \nCodeBERT build a string representation of the buff .   \nBLOCSUM return a string representation of the capability .   \nPython Code def _getcodon_list(codonseq): full_rf_table = codonseq.get_full_rf_table) codon_lst = [] for (i, k) in enumerate(full_rf_table): if isinstance(k, int): start  $=$  k try: end  $=$  int(full_rf_table[(i+1)]) except IndexError: end  $=$  (start + 3) this_codon = str(codonseq[ start:end]) if (len(this_codon) == 3: codon_lst.append(this_codon) else: codon_lst.append(str(this_codon.ungap())) elif (str(codonseq[int(k):(int(k) + 3)) == '---'): codon_lst.append('---') else: codon_lst.append(codonseq[int(k):(int(k) + 3))) return codon_lst   \nGround Truth list of codon accord to full rf table for count .   \nCodeBERT get cod ne .   \nBLOCSUM get list that contain the codon in list .",
+        "guess_lang": "txt",
+        "bbox": [
+            115,
+            118,
+            860,
+            640
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "11439",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? Limitations",
+            "A2. Did you discuss any potential risks of your work? Limitations",
+            "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and 1. Introduction",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            287
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            299,
+            487,
+            316
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "3. Experiment Results",
+        "bbox": [
+            132,
+            321,
+            299,
+            336
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? 3. Experiment Results",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 3. Experiment Results"
+        ],
+        "bbox": [
+            129,
+            346,
+            778,
+            422
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3. Experiment Results and Appendix A"
+        ],
+        "bbox": [
+            127,
+            432,
+            880,
+            753
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "3. Experiment Results and Appendix C",
+        "bbox": [
+            132,
+            785,
+            423,
+            802
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3. Experiment Results and Appendix C",
+        "bbox": [
+            129,
+            812,
+            880,
+            860
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "11440",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            129,
+            84,
+            878,
+            115
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "We used the same hyperparameter as the previous study.",
+        "bbox": [
+            149,
+            117,
+            564,
+            131
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            129,
+            143,
+            880,
+            190
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "We adopted the median value among the 3 models.",
+        "bbox": [
+            149,
+            192,
+            522,
+            206
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            129,
+            217,
+            880,
+            265
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "3. Experiment Results and Appendix B, C",
+        "bbox": [
+            149,
+            267,
+            460,
+            282
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            340,
+            880,
+            372
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            373,
+            349,
+            388
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            399,
+            880,
+            447
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            448,
+            349,
+            463
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            475,
+            880,
+            521
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            523,
+            349,
+            539
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            549,
+            872,
+            565
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            565,
+            349,
+            581
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            592,
+            880,
+            623
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            625,
+            349,
+            640
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "11441",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2023/BLOCSUM_ Block Scope-based Source Code Summarization via Shared Block Representation/bdc012ed-0e12-4395-93e5-cef4ae1613b2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be487ba1edb8533cb6a08b905cb15c8e0a2337dfbdfa8319514446396529c0a8
+size 568595

2023/BLOCSUM_ Block Scope-based Source Code Summarization via Shared Block Representation/full.md

@@ -0,0 +1,426 @@
+# BLOCSUM: Block Scope-based Source Code Summarization via Shared Block Representation
+
+YunSeok Choi, Hyojun Kim, Jee-Hyong Lee
+
+College of Computing and Informatics
+
+Sungkyunkwan University
+
+Suwon, South Korea
+
+{ys.choi, rlagywns0213, john}@skku.edu
+
+# Abstract
+
+Code summarization, which aims to automatically generate natural language descriptions of the source code, has become an essential task in software development for better program understanding. Abstract Syntax Tree (AST), which represents the syntax structure of the source code, is helpful when utilized together with the sequence of code tokens to improve the quality of code summaries. Recent works on code summarization attempted to capture the sequential and structural information of the source code, but they considered less the property that source code consists of multiple code blocks. In this paper, we propose BLOCSUM, BLOck scope-based source Code SUMmarization via shared block representation that utilizes block-scope information by representing various structures of the code block. We propose a shared block position embedding to effectively represent the structure of code blocks and merge both code and AST. Furthermore, we develop variant ASTs to learn rich information such as block and global dependencies of the source code. To prove our approach, we perform experiments on two real-world datasets, the Java dataset and the Python dataset. We demonstrate the effectiveness of BLOCSUM through various experiments, including ablation studies and a human evaluation.
+
+# 1 Introduction
+
+A description of source code is very important in software development because it helps developers better understand programs. Advances in deep learning have enabled automatic code summarization and increased software maintenance efficiency. Previous approaches on automatic source code summarization can be categorized into sequence-based, structure-based, and hybrid approaches. Sequence-based approaches generated summaries by capturing the sequential information of source code (Iyer et al., 2016; Allamanis et al., 2016; Liang and Zhu, 2018; Hu et al., 2018b; Wei et al., 2019; Ye et al.,
+
+2020; Ahmad et al., 2020). They tokenized source code into a sequence of code tokens and encoded them using seq2seq models. Meanwhile, structure-based approaches used Abstract Syntax Tree (AST) to capture the structural information of code (Hu et al., 2018a; Fernandes et al., 2019; Shido et al., 2019; Harer et al., 2019; Zhang et al., 2019; LeClair et al., 2020; Liu et al., 2021; Lin et al., 2021; Allamanis et al., 2018; Wang and Li, 2021; Wu et al., 2021). They parsed the source code into the AST and utilized graph models such as Graph Neural Networks (GNNs). Some works flattened AST into a pre-order traversal sequence (Alon et al., 2019, 2018; LeClair et al., 2019; Wang and Li, 2021; Choi et al., 2021). Hybrid approaches utilized both the token sequence and the ASTs of codes (Wan et al., 2018; Wei et al., 2020; Zhang et al., 2020; Shi et al., 2021). They parallelly processed token sequences and ASTs with independent encoders and tried to merge them in the decoder.
+
+However, the existing approaches have some limitations. Sequence-based approaches treated the source code as a single statement, so they incorporated only the sequence information of the code without any structural information such as code blocks. Structural information is very important to understand code because a snippet of code can be considered as a hierarchy of blocks.
+
+Structure-based approaches tried to catch such structural information of code, but they considered less sequential information of code. A snippet of code is a sequence, so sequential information is also important to understand code. Another problem is that they depended only ASTs to capture structural information of code. However, ASTs are not suitable for capturing structural information because they are syntax trees for grammatical purpose. Since the AST is a tree, there is only one path between every pair of nodes. Any two nodes in ASTs are connected, but with a relatively long path, which hinders capturing the structural rela
+
+![](images/2422250e06caa6d720fdf4d3b70133bba8410b8201061567976c51650d2c4f81.jpg)  
+(a) Source code
+
+![](images/b31eba5f051d5d0451bca247d96cd52f89d3395601994eb0d7d2007145119628.jpg)  
+(b) Abstract Syntax Tree  
+Figure 1: An example of code snippet and its AST in the Java language. (a) The orange box denotes the code block and the number means the index of the code block. (b) The nodes in the orange rectangular box belong to the same block in the AST.
+
+tions of nodes. It causes difficulties in propagating structural information to distant nodes in the AST. Some structure-based approaches tried to alleviate these problems by providing additional graphs such as Control Flow Graph (CFG) and Program Dependence Graph (PDG), but the cost and time required to produce these graphs and integrate them into one graph are not negligible.
+
+Hybrid approaches utilized both code and its AST. Since both sequential and structural information of code is necessary to understand code, the approaches showed higher performances than the previous ones. However, they failed to effectively merge two different types of information. They simply adopted independent encoders for each of them and tried to merge them in the decoder. Due to independent encoders, their representations are easy to be independent. It will make it hard to effectively merge both sequential information of code and structural information. Since the token sequence and the AST of code are just different descriptions of the same code, they need to be encoded so that they are correlated with each other.
+
+To address these limitations mentioned above, we exploit the fact that the source code is a set of blocks consisting of multiple statements for a specific purpose. The code tokens in one code block are configured for the same purpose. As shown in Figure 1a, the code block if  $\{\dots\}$  (orange) consists of statements that are executed when a certain condition is true. So, it needs to consider the information on which block each code token belongs to. To better capture structural information, we need to give each token not only positional
+
+information but also block positional information when encoding.
+
+Since ASTs do not have enough information to capture the structural information of code, we need to modify ASTs. The additional information we try to add is block dependency and global dependency between nodes. For example, as shown in Figure 1b, the node "max" in the orange block of AST is only connected to its parent node, "Statement", but there exists implicit block dependency that the node "max" belongs to the same orange block as the node "data" in the purple dashed line. Furthermore, there exists implicit global dependency between nodes. For example, the node "max" in the orange box is the same variable as the nodes in the green dashed line. We need to add such information to ASTs.
+
+Also, we need to match blocks in the token sequence with blocks in the AST. For example, code tokens ("if", "({", "data", ..., "})" (orange) in Figure 1a can be mapped to its corresponding nodes in the AST ("If", "data", ..., "println") (orange) in Figure 1b. Here, the block in the code and the block in the AST are the same part with an equal role, so we will utilize information on which block of code corresponds to which block of AST. Such information can make the code and AST correlated, and assist effective merging of two kinds of information.
+
+In this paper, we propose BLOCSUM, BLOck scope-based source Code SUMmarization via shared block representation that utilizes blockscope information of token sequences and ASTs. First, we propose the shared block position em
+
+![](images/abe0afa0053a5cf3dc04cdd783ca1d6e7c83350114a2b1e4ddb223c2118ebe8e.jpg)  
+Figure 2: Overview of our proposed approach, BLOCSUM.
+
+beddings for effectively representing the structure of the code block and combining a correlation between the code and the AST encoders. Furthermore, we develop simple yet effective variants of ASTs to learn rich information such as block and global dependencies of the source code. To validate our approach, we perform experiments on the Java dataset and the Python dataset. We prove the superiority of BLOCSUM through various experiments including ablation studies and a human evaluation.
+
+# 2 BLOCSUM
+
+In this section, we present the details of our model. Figure 2 shows the overall architecture of BLOCSUM. We first introduce the shared block position embedding and the abstract syntax tree variants and explain the architecture of BLOCSUM in detail.
+
+# 2.1 Shared Block Position Embedding
+
+We suppose that there are code tokens  $c_{i}$  in a code snippet  $\mathbf{C} = \{c_1, c_2, \ldots\}$  and AST nodes  $n_{i}$  in its AST sequence  $\mathbf{N} = \{n_1, n_2, \ldots\}$ . We aim to predict a summary given the code tokens and the AST nodes.
+
+Code blocks are the basic structural components of source code. Usually, code tokens in a block are gathered for a certain purpose, so the tokens need to be identified that they are in the same block. To distinguish which blocks are, we assign indexes to each block in the order of blocks in the code. Then
+
+each token has a block position with an index of the block it belongs to. If the code token is in nested blocks, we choose the innermost block index as the block position.
+
+In order to utilize the block information of each token, we develop the block position embedding layer. Tokens in the same block have the same block position embedding. The code token embedding for the code encoder,  $E_{c}$ , is defined as follows:
+
+$$
+E _ {c} (t) = W _ {c} (t) + P _ {c} (t) + B _ {c} (t) \tag {1}
+$$
+
+for code token  $t$ . In the equation 1,  $W_{c}$ ,  $P_{c}$ , and  $B_{c}$  are the word, position, and block position embedding layers for the code encoder, respectively. Two position embeddings,  $P_{c}$  and  $B_{c}$ , are learnable positional encoding.
+
+We also combine the AST nodes with block position embeddings to ensure that nodes in the same block have identical block information when node representations are learned by the AST encoder. As with the block position of the code tokens, each node is assigned a block position value. Since the block in the AST is a sub-tree structure in the AST, the node has an index of the sub-tree to which it belongs. Nodes in the same block have the same block positional embedding as code tokens. The AST node embedding for the AST encoder,  $E_{s}$ , is defined as follows:
+
+$$
+E _ {s} (n) = W _ {s} (n) + P _ {s} (n) + B _ {s} (n) \tag {2}
+$$
+
+![](images/3976b24a222534e207b3671b0fe020cadc8f8e2dc90d1d4d7fd593b10db397b2.jpg)  
+(a) AST-original
+
+![](images/92b79d1e047e4112dcd20804f097201b9a35a88eb80f3651f69679d95502c3d3.jpg)  
+(b) AST-block
+
+![](images/5d60ee75777c12d2ca60f0bc860850fad1f6c2196e07b3fbd9abb821e469f8ef.jpg)  
+(c) AST-global  
+Figure 3: We introduce three different types of the AST: (a) AST-original referred as the original AST, (b) AST-block connected by edges between two nodes which are in the same code block, (c) AST-global connected by edges between all the nodes in the AST.
+
+for node  $n$ . In Equation 2,  $W_{s}$ ,  $P_{s}$ , and  $B_{s}$  are the word, position, and block position embedding layers for the AST encoder, respectively. The position of a node is defined as the position in the pre-order traversal sequence of the AST. Two position embeddings,  $P_{s}$  and  $B_{s}$ , are also learnable.
+
+There are two different types of inputs, code and AST, but they are just different descriptions for the same snippet. If two encoders for code and AST learn representations for code tokens and AST nodes separately, their representations will be easy to be independent and very hard to effectively combine both sequential and structural information.
+
+In order to correlate the representations learned by two encoders, we allow the encoders to share the block position embedding layer. If the code token and the AST node belong to the same block, they will have the same block position embedding value. That is, we utilize additional information on which parts of the code correspond to which parts of the AST to generate better representations. If the block position embedding layers are shared, the embeddings for a token,  $t$ , and a node,  $n$  are as follows:
+
+$$
+E _ {c} (t) = W _ {c} (t) + P _ {c} (t) + B (t) \tag {3}
+$$
+
+$$
+E _ {s} (n) = W _ {s} (n) + P _ {s} (n) + B (n) \tag {4}
+$$
+
+where  $B$  is the shared block position embedding layer.
+
+Shared block position embedding can effectively merge the information from two encoders. Also, it helps the code and the AST encoders to capture the structure of source code by providing block information.
+
+# 2.2 Abstract Syntax Tree Variants
+
+The original AST is a structure in which a node is connected only to its parent and children nodes. It contains local information, but it does not include the entire structure information of the code.
+
+Two nodes in the same block have implicit block dependency. There is also global dependency that two nodes have the same meaning even if their hop distance is very long. To utilize rich structural information such as block and global dependencies, we develop a simple yet effective method to reconstruct variants of AST. We define three variants of the AST: AST- original, AST-block, and AST-global.
+
+AST-original is the original AST, which contains information on the local dependency between nodes, as shown in Figure 3a. The variants of AST are graphs of which nodes are the same as those of AST, but of which links are different.
+
+AST-block is the first variant of AST. We obtain it by removing the edges in AST-original, and adding new edges between the nodes belonging to the same block to represent the block structure information, as shown in Figure 3b. It represents information on the block dependency between nodes in the AST.
+
+AST-global is the second variant of AST. As shown in Figure 3c, we fully connect all the nodes in the AST. It represents the global and complete dependency between nodes in the AST. If the pre-order sequence of AST-global is learned by graph model, the node representations in the sequence represent context information of the AST.
+
+Each of the three AST variants represents local, block, and global dependencies between nodes in the AST. When they are learned organically by the AST encoder, node representations will contain rich structural information of the AST.
+
+# 2.3 BLOCSUM Architecture
+
+Code Encoder Our code transformer encoder consists of 6 transformer layers (Vaswani et al., 2017). Each layer of the code transformer encoder is composed of two layers: multi-head self-attention (Vaswani et al., 2017) and feed-forward network. And residual connection (He et al., 2016) and layer normalization (Ba et al., 2016) are performed on each two sub-layers. The transformer encoder captures the sequential and block information of the code tokens to which the shared block position embedding is added.
+
+AST Encoder We use Graph Attention Networks (GATs) (Velickovic et al., 2018) for learning three different AST variants defined above: AST-original, AST-block, and AST-global. Our AST encoder consists of 6 multiple GAT encoder layers. Each layer of the AST multiple GAT encoder consists of three GATs for each variant AST. Each GAT captures
+
+local dependencies for AST-original, block dependencies for AST-block, and global dependencies for AST-global, respectively.
+
+For the  $l$ -th layer of the AST encoder, the process is performed as follows:
+
+$$
+h _ {l n} ^ {l} = G A T _ {l n} \left(A _ {l n}, h _ {n} ^ {l - 1}\right)
+$$
+
+$$
+h _ {b n} ^ {l} = G A T _ {b n} \left(A _ {b n}, h _ {n} ^ {l - 1}\right) \tag {5}
+$$
+
+$$
+h _ {g n} ^ {l} = G A T _ {g n} (A _ {g n}, h _ {n} ^ {l - 1})
+$$
+
+where  $GAT_{ln}, GAT_{bn}, GAT_{gn}$  and  $A_{ln}, A_{bn}, A_{gn}$  denote the GAT layers for three variant ASTs and the adjacency matrices in the AST-original, AST-block, and AST-global, respectively. Especially, the GAT for AST-global is the same as self-attention for learning the context of all nodes in the AST.
+
+Finally, the three representations are combined and performed from residual connection and layer normalization by the following equation:
+
+$$
+h _ {n} ^ {l} = L N \left(h _ {n} ^ {l - 1} + F F N \left(h _ {l n} ^ {l}, h _ {b n} ^ {l}, h _ {g n} ^ {l}\right)\right) \tag {6}
+$$
+
+where  $h_n^l$  is the concatenated node embedding in the  $l$ -th layer of AST GAT encoder,  $LN$  denotes layer normalization, and  $FFN$  is a feed-forward network.
+
+With the deep AST encoder layers, the node representations combine and propagate the local, block, and global information of the AST.
+
+Summary Decoder The summary transformer decoder consists of 6 transformer decoder layers (Vaswani et al., 2017). Code token representations are learned with sequence and block information in the code transformer encoder and node representations are learned with local, block, and global dependencies of the code in the AST encoder. Given the code and node representations learned from each encoder, the summary transformer decoder learns to predict the summary of the original code token by fusion of the code and the AST representations. The multi-head self-attention in the decoder is performed sequentially on the code representations and node representations.
+
+Finally, when the summary transformer decoder predicts the  $t$ -th words, the copy mechanism (See et al., 2017) is applied to directly use the code tokens and AST nodes.
+
+# 3 Experiment Results
+
+# 3.1 Setup
+
+Datasets We evaluate using the benchmarks of the two real-world datasets, the Java dataset (Hu
+
+et al., 2018b) and the Python dataset (Wan et al., 2018). The experiment datasets are divided into 69,708/8,714/8,714 and 55,538/18,505/18,502 for train-valid/test, respectively. For extracting the AST of each dataset, we used a java parser javalang in the Java dataset and a python parser ast in the Python dataset used by Wan et al. (2018). Refer to Appendix A for the statistics of the datasets in detail.
+
+Hyper-parameters We set the maximum length of code, AST, and summary to be 200, 200, and 50, respectively. For training the model, we use Adam optimizer (Kingma and Ba, 2015). We set the minibatch size as 80. The maximum training epoch is 100, and if the performance does not improve for 5 epochs, we stop early. Refer to Appendix C for the implementation details.
+
+Evaluation Metrics We use BLEU (Papineni et al., 2002), METEOR (Banerjee and Lavie, 2005), and ROUGE-L (Lin, 2004) as metrics. We adopt S-BLEU, which indicates the average sentence-level BLEU score. Refer to Appendix B in detail.
+
+Baselines We adopt seq2seq models (Iyer et al., 2016; Hu et al., 2018a,b; Wei et al., 2019; Ahmad et al., 2020), graph2seq models (Eriguchi et al., 2016; Wan et al., 2018; Wu et al., 2021), hybrid models (Choi et al., 2021; Shi et al., 2021; Wu et al., 2021; Gong et al., 2022), and a pre-trained model CodeBERT (Feng et al., 2020) as the baselines. We fine-tuned the pre-trained model CodeBERT for code summarization.
+
+# 3.2 Quantitative Result
+
+Overall Result Table 1 shows the overall performance of BLOCSUM and baselines on the Java and Python benchmark datasets. First, we can see that BLOCSUM improves the performance by 4.47 and 2.64 BLEU, 5.68 and 4.66 METEOR, and 4.63 and 3.85 ROUGE-L on the Java and Python datasets compared to the sequence-based approach, TransRel. In comparison with the structure-based approach, SITTransformer, the performance of BLOCSUM improves by 3.29 and 1.05 BLEU, 4.53 and 2.11 METEOR, and 3.84 and 2.23 ROUGE-L on the two datasets. The result shows that it is effective to capture the overall structure of code when AST is utilized together with the sequence of code tokens. Moreover, BLOCSUM performs better than hybrid approaches. Compared to GCNTransformer, the result shows that it
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">Java</td><td colspan="3">Python</td></tr><tr><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td></tr><tr><td>CODE-NN (Iyer et al., 2016)</td><td>27.60</td><td>12.61</td><td>41.10</td><td>17.36</td><td>09.29</td><td>37.81</td></tr><tr><td>Tree2Seq (Eriguchi et al., 2016)</td><td>37.88</td><td>22.55</td><td>51.50</td><td>20.07</td><td>08.96</td><td>35.64</td></tr><tr><td>RL+Hybrid2Seq (Wan et al., 2018)</td><td>38.22</td><td>22.75</td><td>51.91</td><td>19.28</td><td>09.75</td><td>39.34</td></tr><tr><td>DeepCom (Hu et al., 2018a)</td><td>39.75</td><td>23.06</td><td>52.67</td><td>20.78</td><td>09.98</td><td>37.35</td></tr><tr><td>TL-CodeSum (Hu et al., 2018b)</td><td>41.31</td><td>23.73</td><td>52.25</td><td>15.36</td><td>08.57</td><td>33.65</td></tr><tr><td>Dual Model (Wei et al., 2019)</td><td>42.39</td><td>25.77</td><td>53.61</td><td>21.80</td><td>11.14</td><td>39.45</td></tr><tr><td>Transformer (Ahmad et al., 2020)</td><td>44.58</td><td>26.43</td><td>54.76</td><td>32.52</td><td>19.77</td><td>46.73</td></tr><tr><td>CodeBERT* (Feng et al., 2020)</td><td>41.32</td><td>27.42</td><td>55.33</td><td>30.72</td><td>21.53</td><td>49.93</td></tr><tr><td>GCNTransformer (Choi et al., 2021)</td><td>45.49</td><td>27.17</td><td>54.82</td><td>32.82</td><td>20.12</td><td>46.81</td></tr><tr><td>SiTTransformer (Wu et al., 2021)</td><td>45.76</td><td>27.58</td><td>55.58</td><td>34.11</td><td>21.11</td><td>48.35</td></tr><tr><td>CAST (Shi et al., 2021)</td><td>45.19</td><td>27.88</td><td>55.08</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SCRIPT (Gong et al., 2022)</td><td>46.89</td><td>28.48</td><td>56.69</td><td>34.00</td><td>20.84</td><td>48.15</td></tr><tr><td>BLOCSUM</td><td>49.05</td><td>32.11</td><td>59.42</td><td>35.16</td><td>23.22</td><td>50.58</td></tr></table>
+
+is more effective to capture both the sequential and structural information of the code considering the local, block, and global dependencies of AST rather than a flattened AST. Also, BLOCSUM considering the correlation between code and node representations performs better than Tripletopos using two independent encoders for the code and AST. Finally, we compared our approach with CodeBERT, a strong pre-trained program language model. BLOCSUM performs significantly better than CodeBERT trained on large code data. The result shows that our approach is more appropriate for code modeling than the pre-trained model in the code summarization task.
+
+# 3.3 Qualitative Result
+
+Ablation Study we perform ablation studies to validate the effectiveness of shared block position embedding and AST variants on the Java and Python datasets.
+
+First, we design five models for comparison to verify the shared block position embedding: 1) not use block position embedding (unuse) 2) only use code block position embedding (code block emb) 3) only use AST block position embedding (ast block emb) 4) use separate block position embedding for code and AST (separate) 5) use shared block position embedding (share).
+
+In Table 2, code block emb and ast block emb have better performance than unuse. This result shows that block position embedding is effective in capturing the block information in each encoder. Also when both code and AST encoder use each separate block position embedding, we can see that it is more effective than only using one block po
+
+Table 1: Comparison of our proposed model with the baseline models on the Java and Python datasets. We fine-tuned CodeBERT* with an input length of 200 and an output length of 50 for the two datasets.  
+
+<table><tr><td>Block Pos Emb</td><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td></tr><tr><td colspan="4">Java Dataset</td></tr><tr><td>unused</td><td>48.46</td><td>31.47</td><td>58.65</td></tr><tr><td>code block emb</td><td>48.53</td><td>31.56</td><td>58.63</td></tr><tr><td>ast block emb</td><td>48.54</td><td>31.77</td><td>58.92</td></tr><tr><td>separate</td><td>48.83</td><td>32.06</td><td>59.13</td></tr><tr><td>share</td><td>49.05</td><td>32.11</td><td>59.42</td></tr><tr><td colspan="4">Python Dataset</td></tr><tr><td>unused</td><td>34.33</td><td>22.74</td><td>49.98</td></tr><tr><td>code block emb</td><td>34.50</td><td>22.84</td><td>50.10</td></tr><tr><td>ast block emb</td><td>34.56</td><td>22.83</td><td>50.08</td></tr><tr><td>separate</td><td>34.79</td><td>23.10</td><td>50.33</td></tr><tr><td>share</td><td>35.16</td><td>23.22</td><td>50.58</td></tr></table>
+
+Table 2: Ablation study on shared block position embedding.
+
+sition embedding. Moreover, share has the best performance in comparison with other models. The shared block position embedding can learn the correlation between the code and AST encoders through the block-scope information rather than each separate block position embedding. Shared block position embedding not only effectively captures the structure of the code block, but also helps connect both the code and AST.
+
+Second, we compare our approach with other combinations for verifying the effectiveness of variant ASTs: 1) AST-original (o) 2) AST-block (b), 3) AST-global (g).
+
+As illustrated in Table 3, the results show that leveraging more structural information, such as block or global dependencies, performs better than modeling AST-original with only local dependency. Also, combining AST-original, AST-block, and AST-global has the best performance in comparison with
+
+<table><tr><td>Combination</td><td>BLEU</td><td>METEOR</td><td>ROUGE-L</td></tr><tr><td colspan="4">Java Dataset</td></tr><tr><td>o</td><td>48.39</td><td>31.73</td><td>58.98</td></tr><tr><td>g</td><td>48.41</td><td>31.63</td><td>58.40</td></tr><tr><td>o + g</td><td>48.76</td><td>31.87</td><td>59.00</td></tr><tr><td>o + b + g</td><td>49.05</td><td>32.11</td><td>59.42</td></tr><tr><td colspan="4">Python Dataset</td></tr><tr><td>o</td><td>34.58</td><td>22.76</td><td>50.14</td></tr><tr><td>g</td><td>34.41</td><td>22.84</td><td>50.01</td></tr><tr><td>o + g</td><td>34.64</td><td>22.87</td><td>50.11</td></tr><tr><td>o + b + g</td><td>35.16</td><td>23.22</td><td>50.58</td></tr></table>
+
+combining other combinations. We demonstrate that utilizing AST variants helps learn rich information such as block and global dependencies of the source code.
+
+Human Evaluation We performed a human evaluation on the Python dataset to demonstrate the quality of generated summaries. We randomly select 100 code snippets and ask three people with knowledge of the Python language to evaluate the summaries. They are CS graduate students with many years of experience in Python languages. Following the human evaluation metrics of (Choi et al., 2021), we ask them to evaluate the 3 following metrics: 1) Fluency (Quality of the summary & grammatically correct), 2) Relevance (Selection of the consistent content in source code), 3) Coverage (Selection of the important content in source code). We show pairs of summaries generated from BLOCSUM and the baseline fine-tuned (Feng et al., 2020) to the evaluators, and they select one of win, tie, and loss in three metrics for both results.
+
+Table 4 shows the results of human evaluation on the generated summaries on the Python dataset. The scores of fluency is lower, but the relevance and coverage are very higher than the baseline, CodeBERT. We analyzed the generated summaries of the two models and identified that BLOCSUM generates it similarly to the ground truth, reflecting the keyword of the code. CodeBERT, a pretrained language model, can generate more fluent and grammatical summaries, but the length is relatively short and a very plain summary with no keywords. The average tokens in the ground truth are 10.14, while the average tokens in summaries generated by BLOCSUM and CodeBERT are 9.91 and 8.16, respectively. We think that short sentences are more grammatically advantageous than long sentences. BLOCSUM has the highest tie in terms
+
+Table 3: Ablation study on combination of AST variants.  
+
+<table><tr><td></td><td>Fluency</td><td>Relevance</td><td>Coverage</td></tr><tr><td>Win</td><td>76</td><td>143</td><td>125</td></tr><tr><td>Tie</td><td>114</td><td>64</td><td>80</td></tr><tr><td>Loss</td><td>110</td><td>93</td><td>95</td></tr></table>
+
+Table 4: Human evaluation of the appropriateness of the generated adversarial example on the Python dataset.
+
+of fluency but the highest win in terms of relevance and coverage. The result means that BLOCSUM reflects more the core characteristic of the code than CodeBERT.
+
+Comparison with the baselines Table 5 shows the summary examples generated from our proposed model on the Python dataset. The result on the Python dataset example shows BLOCSUM predicts the keywords "wsgi" and "request" by reflecting block-scope information. Although there is no dependency between two words in the code and the original AST, BLOCSUM utilizes three different types of AST variants and jointly learns structural dependency in three aspects to improve the performance of the model. Refer to Appendix D for more summary examples on the Java and Python datasets.
+
+# 4 Related Work
+
+Sequence-based Approaches Iyer et al. (2016) and Allamanis et al. (2016) proposed to use Long Short Term Memory (LSTM) and Convolutional Neural Networks (CNNs) for the source code summarization. Liang and Zhu (2018) proposed a tree-based recursive neural network to represent the syntax tree of code. Hu et al. (2018b) and Chen et al. (2021) summarized the source code with the APIs knowledge. Wei et al. (2019) used a dual training framework by training code summarization and code generation tasks. Also, Ye et al. (2020) considered the probabilistic correlation between the two tasks. Choi et al. (2020) proposed attention-based keyword memory networks for code summarization. Ahmad et al. (2020) proposed a Transformer model using a relative position. However, these approaches have limitations in that they did not explicitly incorporate the structural information of the source code, which is just as crucial as capturing the code semantics. Also, they did not learn the code block information because they learned the code as a sequence of tokens.
+
+<table><tr><td>Python Code</td><td>def simulate_request(app, method=&#x27;GET&#x27;, path=&#x27;/&#x27;, query_string=None, headers=None, body=None, fileRIAWRER=None, params=None, params_csv=True): if (not path.startswith(&#x27;/&#x27;)) : raise ValueError(&quot;path must start with &#x27;/&quot;) if (query_string and query_string.startswith(&#x27;?&#x27;)) : raise ValueError(&quot;query_string should not start with &#x27;?&#x27;) if (&#x27;?&#x27; in path): raise ValueError(&#x27;path may not contain a query string. Please use the query_string parameter instead.&#x27;) if (query_string is None): query_string = to_query_str(parms, comma delimitedLists=params_csv, prefix=False) env = helpers.create_environ(method=method, path=path, query_string=(query_string or ), headers=headers, body=body, file wraps=file wraps) srmock = StartResponseMock(   ) validator = wsgirefvalidate_validator(app) iterable = validator(env, smock) result = Result(iterable, smock.status, smockHeaders) return result</td></tr><tr><td>Ground Truth CodeBERT BLOCSUM</td><td>simulate a request to a wsgi application . simulate a wsgi environment . simulate a request to a wsgi request .</td></tr></table>
+
+Table 5: A qualitative example on the Python dataset.
+
+Structure-based Approaches Hu et al. (2018a) proposed an RNN-based model using the pre-order traversal sequence as input. Shido et al. (2019); Harer et al. (2019) adopted Tree-LSTM, TreeTransformer to encode tree-based inputs. LeClair et al. (2020) proposed encoded AST using graph neural networks and trained LSTM. Liu et al. (2021) proposed a retrieval augmented method with Graph Neural Network (GNN). Zhang et al. (2019) proposed AST-based Neural Network (ASTNN) for encoding the subtree. Lin et al. (2021) proposed Tree-LSTM to represent the split AST for code summarization. Li et al. (2021) leverage the retrieve-and-edit framework to improve the performance for code summarization. Allamanis et al. (2018), Wang and Li (2021), and Wu et al. (2021) tried to capture rich information using additional graphs such as CFG and PDG. But, these approaches considered only the structural information of the AST without considering the sequential information of the code token.
+
+Hybrid Approaches Alon et al. (2019) leveraged the unique syntactic structure of programming languages by sampling paths in the AST of a code snippet. LeClair et al. (2019) proposed ast-attendgru model that combines code with structure from AST Also, Choi et al. (2021) proposed a model that combines Graph Convolution Network and Transformer using AST. Wu et al. (2021) incorporated a multiview graph matrix into the transformer model. Shi et al. (2021) tried to hierarchically split and
+
+reconstruct ASTs using Recursive Neural Network for learning the representation of the complete AST. Wan et al. (2018) used a deep reinforcement learning framework to consider an AST structure and code snippets. Gong et al. (2022) proposed a structural position method to augment the structural correlations between code tokens. But, they encoded two types of representations independently without correlation and did not consider merging them.
+
+Zhang et al. (2020) proposed a retrieval-based approach using syntactic and semantic similarity for source code summarization. Liu et al. (2021) proposed a hybrid GNN using a retrieval augmented graph method. Wei et al. (2020) proposed a comment generation framework using AST, similar code, and exemplar from code. Choi et al. (2023) proposed a self-attention network that adaptively learns the structural and sequential information of code. But, they tried to model the code using more code information through retrieval methods.
+
+# 5 Conclusion
+
+In this paper, we proposed BLOCSUM, BLOck scope-based source Code SUMmarization via shared block representation that utilizes blockscope information by representing various structures of the code block. We designed two methods using the fact that a code block is a fundamental structural component of the source code. We propose the first method, the shared block position embedding, for effectively representing the structure of the code block and merging a correlation
+
+between the code and the AST encoders. Furthermore, we developed to reconstruct simple yet effective AST variants to learn rich information such as block and global dependencies of the source code. Experimental results demonstrated the effectiveness of BLOCSUM and confirmed the importance of block-scope information in the code.
+
+# Limitations
+
+In this paper, we conducted an experiment on code summarization using two benchmark datasets, the Java dataset (Hu et al., 2018b) and the Python dataset (Wan et al., 2018). BLOCSUM may need to be tested for its generalizability to other program languages. We chose two program languages (Java and Python) that were easily parsed to map the block position of Code and AST. We believe that since other programming languages have similar syntactic structures, BLOCSUM should be able to achieve similar performance on them as well.
+
+# Ethics Statement
+
+This paper proposes block scope-based source code summarization via shared block representation that utilizes block-scope information by representing various structures of the code block, which is beneficial to increase the efficiency of developers. The research conducted in this paper will not cause any ethical issues or have any negative social effects. The data used is all publicly accessible and is commonly used by researchers as a benchmark for program and language generation tasks. Our proposed method does not introduce any ethical or social bias or worsen any existing bias in the data.
+
+# Acknowledgements
+
+This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No.2019-0-00421, AI Graduate School Support Program(Sungkyunkwan University)) This research was supported by the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience Program(IITP-2023-2020-0-01821) supervised by the IITP(Institute for Information & communications Technology Planning & Evaluation)
+
+# References
+
+Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. 2020. A transformer-based approach for source code summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4998-5007, Online. Association for Computational Linguistics.  
+Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. 2018. Learning to represent programs with graphs. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.  
+Miltiadis Allamanis, Hao Peng, and Charles Sutton. 2016. A convolutional attention network for extreme summarization of source code. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 2091-2100. JMLR.org.  
+Uri Alon, Shaked Brody, Omer Levy, and Eran Yahav. 2019. code2seq: Generating sequences from structured representations of code. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. 2018. code2vec: Learning distributed representations of code.  
+Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normalization. *ArXiv preprint*, abs/1607.06450.  
+Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, Ann Arbor, Michigan. Association for Computational Linguistics.  
+Fuxiang Chen, Mijung Kim, and Jaegul Choo. 2021. Novel natural language summarization of program code via leveraging multiple input representations. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2510-2520, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+YunSeok Choi, JinYeong Bak, CheolWon Na, and JeeHyong Lee. 2021. Learning sequential and structural information for source code summarization. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 2842–2851, Online. Association for Computational Linguistics.  
+YunSeok Choi, Suah Kim, and Jee-Hyong Lee. 2020. Source code summarization using attention-based
+
+keyword memory networks. In 2020 IEEE International Conference on Big Data and Smart Computing (BigComp), pages 564-570. IEEE.  
+YunSeok Choi, CheolWon Na, Hyojun Kim, and JeeHyong Lee. 2023. Readsum: Retrieval-augmented adaptive transformer for source code summarization. IEEE Access.  
+Akiko Eriguchi, Kazuma Hashimoto, and Yoshimasa Tsuruoka. 2016. Tree-to-sequence attentional neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 823-833, Berlin, Germany. Association for Computational Linguistics.  
+Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A pre-trained model for programming and natural languages. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1536-1547, Online. Association for Computational Linguistics.  
+Patrick Fernandes, Miltiadis Allamanis, and Marc Brockschmidt. 2019. Structured neural summarization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Zi Gong, Cuiyun Gao, Yasheng Wang, Wenchao Gu, Yun Peng, and Zenglin Xu. 2022. Source code summarization with structural relative position guided transformer. arXiv preprint arXiv:2202.06521.  
+Jacob Harer, Chris Reale, and Peter Chin. 2019. Tree-transformer: A transformer-based method for correction of tree-structured data. ArXiv preprint, abs/1908.00449.  
+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.  
+Xing Hu, Ge Li, Xin Xia, David Lo, and Zhi Jin. 2018a. Deep code comment generation. In 2018 IEEE/ACM 26th International Conference on Program Comprehension (ICPC), pages 200-20010. IEEE.  
+Xing Hu, Ge Li, Xin Xia, David Lo, Shuai Lu, and Zhi Jin. 2018b. Summarizing source code with transferred API knowledge. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden, pages 2269-2275. ijcai.org.  
+Srinivasan Iyer, Ioannis Konstas, Alvin Cheung, and Luke Zettlemoyer. 2016. Summarizing source code using a neural attention model. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages
+
+2073-2083, Berlin, Germany. Association for Computational Linguistics.  
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Alexander LeClair, Sakib Haque, Lingfei Wu, and Collin McMillan. 2020. Improved code summarization via a graph neural network. volume abs/2004.02843.  
+Alexander LeClair, Siyuan Jiang, and Collin McMillan. 2019. A neural model for generating natural language summaries of program subroutines. In 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE), pages 795-806. IEEE.  
+Jia Li, Yongmin Li, Ge Li, Xing Hu, Xin Xia, and Zhi Jin. 2021. Editsum: A retrieve-and-edit framework for source code summarization. In 2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE), pages 155-166. IEEE.  
+Yuding Liang and Kenny Qili Zhu. 2018. Automatic generation of text descriptive comments for code blocks. 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 5229-5236. AAAI Press.  
+Chen Lin, Zhichao Ouyang, Junqing Zhuang, Jianqiang Chen, Hui Li, and Rongxin Wu. 2021. Improving code summarization with block-wise abstract syntax tree splitting. In 2021 IEEE/ACM 29th International Conference on Program Comprehension (ICPC), pages 184-195. IEEE.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Shangqing Liu, Yu Chen, Xiaofei Xie, Jing Kai Siow, and Yang Liu. 2021. Retrieval-augmented generation for code summarization via hybrid GNN. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational
+
+Linguistics (Volume 1: Long Papers), pages 1073-1083, Vancouver, Canada. Association for Computational Linguistics.  
+Ensheng Shi, Yanlin Wang, Lun Du, Hongyu Zhang, Shi Han, Dongmei Zhang, and Hongbin Sun. 2021. CAST: Enhancing code summarization with hierarchical splitting and reconstruction of abstract syntax trees. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4053-4062, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yusuke Shido, Yasuaki Kobayashi, Akihiro Yamamoto, Atsushi Miyamoto, and Tadayuki Matsumura. 2019. Automatic source code summarization with extended tree-lstm. volume abs/1906.08094.  
+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 Liò, and Yoshua Bengio. 2018. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.  
+Yao Wan, Zhou Zhao, Min Yang, Guandong Xu, Haochao Ying, Jian Wu, and Philip S Yu. 2018. Improving automatic source code summarization via deep reinforcement learning. In Proceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering, pages 397-407.  
+Yanlin Wang and Hui Li. 2021. Code completion by modeling flattened abstract syntax trees as graphs. In AAAI.  
+Bolin Wei, Ge Li, Xin Xia, Zhiyi Fu, and Zhi Jin. 2019. Code generation as a dual task of code summarization. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 6559-6569.  
+Bolin Wei, Yongmin Li, Ge Li, Xin Xia, and Zhi Jin. 2020. Retrieve and refine: exemplar-based neural comment generation. In 2020 35th IEEE/ACM International Conference on Automated Software Engineering (ASE), pages 349-360. IEEE.  
+Hongqiu Wu, Hai Zhao, and Min Zhang. 2021. Code summarization with structure-induced transformer. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 1078-1090, Online. Association for Computational Linguistics.
+
+Wei Ye, Rui Xie, Jinglei Zhang, Tianxiang Hu, Xiaoyin Wang, and Shikun Zhang. 2020. Leveraging code generation to improve code retrieval and summarization via dual learning. In WWW '20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, pages 2309-2319. ACM / IW3C2.  
+Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, and Xudong Liu. 2020. Retrieval-based neural source code summarization. In 2020 IEEE/ACM 42nd International Conference on Software Engineering (ICSE), pages 1385-1397. IEEE.  
+Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, Kaixuan Wang, and Xudong Liu. 2019. A novel neural source code representation based on abstract syntax tree. In 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE), pages 783-794. IEEE.
+
+# A Statistics of Experiment Datasets
+
+For obtaining ASTs of the Java and Python dataset, we use the javalang<sup>1</sup> and ast<sup>2</sup> library, respectively. Also, we tokenize the source code and the AST to subtokens as the form CamelCase and snake-case.
+
+<table><tr><td>Dataset</td><td>Java</td><td>Python</td></tr><tr><td>Train</td><td>69,708</td><td>55,538</td></tr><tr><td>Valid</td><td>8,714</td><td>18,505</td></tr><tr><td>Test</td><td>8,714</td><td>18,502</td></tr><tr><td>Unique non-leaf nodes in ASTs</td><td>106</td><td>54</td></tr><tr><td>Unique leaf nodes in ASTs</td><td>57,372</td><td>101,229</td></tr><tr><td>Unique tokens in summaries</td><td>46,895</td><td>56,189</td></tr><tr><td>Avg. nodes in AST</td><td>131.72</td><td>104.11</td></tr><tr><td>Avg. tokens in summary</td><td>17.73</td><td>9.48</td></tr></table>
+
+Table 6: Statistics of Java dataset (Hu et al., 2018b) and Python dataset (Wan et al., 2018). For obtaining their corresponding ASTs, we use the javalang and ast libraries, respectively.
+
+# B Evaluation Metrics
+
+BLEU(Papineni et al., 2002) is a Bilingual Evaluation Understudy to measure the quality of generated code summaries. The formula for computing BLEU is as follows:
+
+$$
+B L E U = B P \cdot \exp \sum_ {n = 1} ^ {N} \omega_ {n} \log p _ {n}
+$$
+
+where  $p_n$  is the geometric average of the modified n-gram precisions,  $\omega_{n}$  is uniform weights  $1 / N$  and BP is the brevity penalty.
+
+METEOR(Banerjee and Lavie, 2005) is used to measure how closely the metric scores match the human judgments about the quality of the translation. So unigram precision  $(P)$  and unigram recall  $(R)$  are computed and combined via a harmonic mean. The METEOR score is computed as follows:
+
+$$
+M E T E O R = (1 - \gamma \cdot f r a g ^ {\beta}) \cdot \frac {P \cdot R}{\alpha \cdot P + (1 - \alpha) \cdot P}
+$$
+
+where  $\text{frag}$  is the fragmentation fraction.  $\alpha$ ,  $\beta$ , and  $\gamma$  are three penalty parameters whose default values are 0.9, 3.0, and 0.5, respectively.
+
+ROUGE-L(Lin, 2004) is used to apply Longest Common Subsequence in summarization evaluation. ROUGE-L used LCS-based F-measure to estimate the similarity between two summaries  $X$
+
+of length  $m$  and  $Y$  of length  $n$ , assuming  $X$  is a reference summary sentence, and  $Y$  is a candidate summary sentence, as follows:
+
+$$
+\begin{array}{l} R _ {l c s} = \frac {L C S (X , Y)}{m}, P _ {l c s} = \frac {L C S (X , Y)}{n} \\ F _ {l c s} = \frac {\left(1 + \beta^ {2}\right) R _ {l c s} P _ {l c s}}{R _ {l c s} + \beta^ {2} P _ {l c s}} \\ \end{array}
+$$
+
+where  $\beta = P_{lcs} / R_{lcs}$  and  $F_{lcs}$  is the value of ROUGE-L.
+
+# C Implementation Detail
+
+We conducted experiments on Ubuntu 18.04 with 4 2080 Ti GPUs. The environment of the sever supports python 3.9, Cuda 10.2, pytorch 1.9, and pytorch geometric 1.7.
+
+The average training and inference time for BLOCSUM takes about 40 and 0.5 hours, respectively. BLOCSUM has about 76 million parameters.
+
+<table><tr><td>Hyper-parameter</td><td>Size</td></tr><tr><td>the maximum length of code tokens</td><td>200</td></tr><tr><td>the maximum length of AST nodes</td><td>200</td></tr><tr><td>the maximum length of summary tokens</td><td>50</td></tr><tr><td>Embedding dimension</td><td>512</td></tr><tr><td>The number of Code Encoder layers</td><td>6</td></tr><tr><td>The number of AST Encoder layers</td><td>6</td></tr><tr><td>The number of Code Decoder layers</td><td>6</td></tr><tr><td>Head of Attention</td><td>8</td></tr><tr><td>batch size</td><td>80</td></tr><tr><td>train epoch</td><td>100</td></tr><tr><td>early stop</td><td>5</td></tr><tr><td>learning rate</td><td>0.0005</td></tr><tr><td>learning decay</td><td>0.99</td></tr><tr><td>beam size</td><td>5</td></tr></table>
+
+Table 7: Hyper-parameters of BLOCSUM.
+
+# D Examples of Java and Python Datasets
+
+```txt
+Java Code @ Override public String toString() { String result; result=super.toString(); if (m_CapabilitiesFilter != null) { initCapabilities(); if (m_Capabilities != null) { if (m_CapabilitiessupportsMaybe(m_CapabilitiesFilter) &&! m_Capabilitiessupports(m_CapabilitiesFilter)) { result="<html><font color  $\equiv$  "" + MAYBE.'<sup>+</sup> + result + </font></i><html>"; } else if(!m_Capabilitiessupports(m_CapabilitiesFilter)) { result="<html><font color  $\equiv$  "" + NO.'<sup>+</sup> + result+ </font></i><html>"; } } return result;   
+}   
+Ground Truth return a string representation of this tree node .   
+CodeBERT build a string representation of the buff .   
+BLOCSUM return a string representation of the capability .   
+Python Code def _getcodon_list(codonseq): full_rf_table = codonseq.get_full_rf_table) codon_lst = [] for (i, k) in enumerate(full_rf_table): if isinstance(k, int): start  $=$  k try: end  $=$  int(full_rf_table[(i+1)]) except IndexError: end  $=$  (start + 3) this_codon = str(codonseq[ start:end]) if (len(this_codon) == 3: codon_lst.append(this_codon) else: codon_lst.append(str(this_codon.ungap())) elif (str(codonseq[int(k):(int(k) + 3)) == '---'): codon_lst.append('---') else: codon_lst.append(codonseq[int(k):(int(k) + 3))) return codon_lst   
+Ground Truth list of codon accord to full rf table for count .   
+CodeBERT get cod ne .   
+BLOCSUM get list that contain the codon in list .
+```
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? Limitations  
+A2. Did you discuss any potential risks of your work? Limitations  
+A3. Do the abstract and introduction summarize the paper's main claims? Abstract and 1. Introduction  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+3. Experiment Results
+
+B1. Did you cite the creators of artifacts you used? 3. Experiment Results  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 3. Experiment Results
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3. Experiment Results and Appendix A
+
+C Did you run computational experiments?
+
+3. Experiment Results and Appendix C
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3. Experiment Results and Appendix C
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+We used the same hyperparameter as the previous study.
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+We adopted the median value among the 3 models.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+3. Experiment Results and Appendix B, C
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+Not applicable. Left blank.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+Not applicable. Left blank.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+Not applicable. Left blank.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+Not applicable. Left blank.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+Not applicable. Left blank.

2023/BLOCSUM_ Block Scope-based Source Code Summarization via Shared Block Representation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56047c5ecc5733662d502bd13d55bfb566c94dfcf8dc339bbc0a44e6a9eb3b78
+size 577577

2023/Balancing the Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer/e52f8ce0-4325-4d7a-9b9b-68ca012cb308_content_list.json

@@ -0,0 +1,1767 @@
+[
+    {
+        "type": "text",
+        "text": "Balancing Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer",
+        "text_level": 1,
+        "bbox": [
+            132,
+            89,
+            867,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Debarati Das David Ma Dongyeop Kang",
+        "bbox": [
+            305,
+            155,
+            697,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science, University of Minnesota",
+        "bbox": [
+            260,
+            173,
+            739,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{das00015, maxxx818, dongyeop}@umn.edu",
+        "bbox": [
+            305,
+            189,
+            697,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Text style transfer is an exciting task within the field of natural language generation that is often plagued by the need for high-quality paired datasets. Furthermore, training a model for multi-attribute text style transfer requires datasets with sufficient support across all combinations of the considered stylistic attributes, adding to the challenges of training a style transfer model. This paper explores the impact of training data input diversity on the quality of the generated text from the multi-style transfer model. We construct a pseudo-parallel dataset by devising heuristics to adjust the style distribution in the training samples. We balance our training dataset using marginal and joint distributions to train our style transfer models. We observe that a balanced dataset produces more effective control effects over multiple styles than an imbalanced or skewed one. Through quantitative analysis, we explore the impact of multiple style distributions in training data on style-transferred output. These findings will better inform the design of style-transfer datasets.",
+        "bbox": [
+            141,
+            280,
+            460,
+            621
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            634,
+            258,
+            651
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multi-style text transfer is a challenging task today with applications such as automatic domain-appropriate, style-conformant writing (Fu et al., 2018) and AI-assisted stylistic language editing. Text style transfer is an intricate task as all language has a specific context, and those contexts influence the attributes of the language (Hovy and Yang, 2021). Text style transfer is challenging because it involves dealing with the aspects of style coupled with the textual content (Hu et al., 2017; Shen et al., 2017; Lample et al., 2018). This domain's other obstacles include the need for parallel corpus (Jhamtani et al., 2017) and quality training data. As the number of style dimensions increases with multi-style text transfer, not only is the requirement of a jointly annotated corpus across all the",
+        "bbox": [
+            112,
+            661,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2794bbad8e426aa268b19892248a0e6d7b2378e8b55fe42f584182a0884f29f0.jpg",
+        "image_caption": [
+            "Figure 1: When an input sentence is passed to the multi-style transfer model, to increase formality and decrease arousal, we hypothesize that when the model is trained on a balanced joint distribution of formality and arousal (all four style combinations have a  $25\\%$  representation) - the style transfer is more successful as opposed to when the model is trained on a skewed joint distribution (there is no representation of the \"informal unaroused\" style combination) of styles in the training data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            536,
+            255,
+            884,
+            394
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "stylistic dimensions problematic, but the different styles are not necessarily independent.",
+        "bbox": [
+            507,
+            577,
+            880,
+            609
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While \"style\" can also refer to authorial or domain-specific style, in this paper, we focus on \"microstyles\" as defined by (Kang and Hovy, 2021) where they define \"micro-style\" as a complex combination of different factors such as formality markers, emotions, and metaphors. People intentionally (Troiano et al., 2021) tune these styles in writing differently based on their mood, the person they are addressing, the content of the message, or the platform. Multiple microstyles can jointly describe a text; for example, a given text could simultaneously be formal and sad. Microstyles also more easily lend themselves to being represented as spectra with varying degrees of intensity. These points align with our vision of an application where users can edit micro-style aspects of their writing.",
+        "bbox": [
+            507,
+            611,
+            882,
+            868
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Much research exists on models implementing multi-style text transfer and interdependency of micro-styles (Kang and Hovy, 2019; Goyal et al.,",
+        "bbox": [
+            507,
+            870,
+            882,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3932",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 3932-3943",
+        "bbox": [
+            228,
+            945,
+            768,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            958,
+            699,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8fb69b8365f1aa2b8a81111b213cfabf68d9acec2e6d93d651dc5ce7fa6ff820.jpg",
+        "image_caption": [
+            "Figure 2: The input sentence transitions through every step in our multi-style text style transfer pipeline. The box in red indicates our main contribution to the pipeline, which helps us explore the effects of joint micro-style combinations on style-transferred output."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            196,
+            85,
+            801,
+            303
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2020; Subramanian et al., 2018). However, there needs to be more exploration of the joint distribution of inherent micro-styles in the style transfer training dataset and how these micro-style distributions are related. Therefore, we pose a question - Can a dataset with minimal variance across multiple micro-style combinations, such that it experiences a \"balancing effect\", lead to a better style transferred output? Figure 1 illustrates our intuition that a dataset that experiences a \"balancing effect\" will have more control over the multi-style transferred output than a \"skewed\" dataset. Suppose the style transfer model sees examples of every style combination that can exist - this could aid in the style generation of even unlikely combinations of styles compared to a skewed distribution of these joint micro-styles.",
+        "bbox": [
+            115,
+            382,
+            489,
+            657
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this research, we consider a multi-style text style transfer pipeline assuming that the user has no access to parallel data or the style of the original text that he wishes to transfer, as would seem natural for a style language editing application. We introduce the changing of the training dataset micro-style joint distributions in such a pipeline and quantitatively explore the impact of this modification on the style transferred output. We perform a set of empirical analyses to demonstrate the influence of joint distributions on style-transferred output and show how this trend varies as the number of micro-styles considered changes. The 'balancing effect' on a training dataset leads to style transferred sentences from even the joint style combinations that are typically rare (\"informal unbiased and",
+        "bbox": [
+            112,
+            661,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "unaroused\"). Our study is the first of its kind on the distribution of micro styles in training datasets for multi-style text style transfer and is likely to have implications for designing datasets for multi-style transfer model training and fall within the context of and align with recent work on characterizing datasets and factors impacting style transfer (Bender and Friedman, 2018; Schoch et al., 2021; Li et al., 2019; Zhang et al., 2020; Gururangan et al., 2018).",
+        "bbox": [
+            507,
+            382,
+            884,
+            544
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Multi Style Transfer Pipeline",
+        "text_level": 1,
+        "bbox": [
+            507,
+            556,
+            794,
+            571
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Datasets: We chose four micro-styles from the style hierarchy defined in Troiano et al.: Formality, Arousal, Sentiment, and Bias, for our study and used publicly available NLP datasets built by other researchers (Rao and Tetreault, 2018; Buechel and Hahn, 2022; Go et al., 2009; Pryzant et al., 2020; Kang and Hovy, 2019) to develop and test our models. Appendix A mentions the details of the datasets and their usage.",
+        "bbox": [
+            507,
+            581,
+            882,
+            725
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pipeline Overview: Our experimental setup for multi-style transfer is inspired by the work of (Krishna et al., 2020). Like them, we first generate a \"diverse\" paraphrase of the input sentence, and then the paraphrased sentence is rewritten in the style of choice. Towards this end, we train a paraphrasing model (separately on a parallel paraphrase dataset). Then, the trained paraphrase model is used to create \"pseudo-gold\" parallel data for training style models.",
+        "bbox": [
+            505,
+            726,
+            884,
+            885
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "First, we adopted a pre-trained T5 model (Raffel et al., 2020) to generate paraphrases. This model",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3933",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "was trained for the task of paraphrase generation on the ParaNMT-filtered dataset provided by (Krishna et al., 2020). Once we had this trained paraphrase model, we used diverse beam search (Vijayakumar et al., 2016) to generate diverse fluent paraphrased outputs. An important assumption is that the paraphrase is stripped of its original style and does not leak into the training.",
+        "bbox": [
+            112,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We address this potential issue by training classifiers (Sanh et al., 2019) to predict style on the original and paraphrased datasets and find that all our micro-style classifiers have a classification accuracy of higher than  $80\\%$  F1, which is acceptable for pseudo-label creation. After we generate diverse paraphrases, we choose the most diverse paraphrase and then derive micro-style classifications for the paraphrased sentence using our trained micro-style classifiers. Therefore each sentence is assigned a classification score for each micro-style label and can form a \"pseudo parallel\" dataset for training the T5-based joint transfer model. Thus, our approach does not need a parallel dataset.",
+        "bbox": [
+            115,
+            217,
+            489,
+            443
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We then converted the classifier predictions into buckets of style (ranging from \"very low\" to \"very high\") based on the chosen style of the original and then paraphrased sentences. The bucketing process is described in Appendix B. After this step, we introduce our contribution of \"constructing style distributions\" into the pipeline, as illustrated in Figure 2. Following that, we perform multi-style text style transfer. We appended the \"bucket\" information to the paraphrased sentence to achieve the necessary intensity transfers, as motivated by the original T5 paper (Raffel et al., 2020). We train T5-based style transfer models, where the paraphrased sentence and its style buckets are used as input parameters, while the style buckets assigned to the anchor sentence are used as proxy levels of output style transfer. All model-specific details are provided in Appendix B. For generating sentences from our trained models, we used beam search (Vijayakumar et al., 2016) and nucleus sampling (Holtzman et al., 2019) and chose the top 3 sentences from the generations. The following is an example of the input to the joint style transfer model and the expected output.",
+        "bbox": [
+            115,
+            448,
+            489,
+            834
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Goal - Highly increase the formality of the sentence, slightly increase the arousal of the sentence",
+        "bbox": [
+            112,
+            854,
+            487,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Input - transfer: I'm sad you're going | input",
+        "bbox": [
+            112,
+            903,
+            487,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "formality: low | input arousal: low | output formality: high | output arousal: mid Output - I am sorry you are going to go.",
+        "bbox": [
+            507,
+            84,
+            882,
+            133
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Thus, we implemented a multi-style transfer pipeline to test our hypothesis without any finicky modeling paradigms popular in style transfer research, such as variational inference or autoregressive sampling (He et al., 2020; Subramanian et al., 2018).",
+        "bbox": [
+            507,
+            148,
+            884,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/dfdde98d5d8560c139e427a8ce9d1f9665065c0ac37a7a82be3df80fc0245e10.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Style Combination</td><td>Balanced</td><td>Skewed</td></tr><tr><td>Formal Aroused</td><td>3395</td><td>8685</td></tr><tr><td>Formal Unaroused</td><td>3395</td><td>2792</td></tr><tr><td>Informal Aroused</td><td>3395</td><td>1275</td></tr><tr><td>Informal Unaroused</td><td>3395</td><td>828</td></tr></table>",
+        "bbox": [
+            521,
+            256,
+            873,
+            350
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Training data statistics (number of samples) for the balanced and skewed settings, when considering the micro-styles of Formality and Arousal.",
+        "bbox": [
+            507,
+            359,
+            882,
+            401
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Constructing Micro-style Distributions We define a \"style combination\" as a possible combination of the states that the micro-styles can take together - such as 'informal biased negative.' Since there are three micro-styles, each having binary states, the total possible number of style combinations, in this case, is given by  $N_{c} = 2 \\times 2 \\times 2 = 2^{3}$ . Therefore to generalize, if  $|m_{i}|$  indicates the cardinality of each micro-style and  $n$  indicates the number of micro-styles considered, the total possible number of style combinations  $(N_{c})$  possible is given by:",
+        "bbox": [
+            507,
+            422,
+            884,
+            613
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nN _ {c} = \\prod_ {i = 1} ^ {n} | m _ {i} | \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            640,
+            615,
+            882,
+            655
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To create the balanced joint distribution of styles, we ensure the standard deviation across the style combinations is close to 0. We do this by down-sampling each style combination, such that the number of samples in each style combination is the same as the least represented style combination. As we increase micro-styles, some micro-style combinations do not occur naturally together, so their representation is close to 0. In such cases, we assume that the least represented style combination is at least  $5\\%$  of the total dataset. To ensure our comparison across the \"balanced\" and \"skew\" settings is fair, we construct a skewed dataset with a total sample size that is the same as that of the balanced dataset. Thus, the balanced dataset has a uniform distribution, while the skewed dataset",
+        "bbox": [
+            507,
+            661,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3934",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "has a non-uniform distribution. Table 1 shows the number of samples in each style combination of Formality and Arousal, given a \"balanced\" and \"skewed\" setting.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Experimental Results and Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            164,
+            467,
+            181
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics: Style transfer accuracy metrics quantify how nicely output texts match the desired style. However, more than this metric is required. Motivated by Jin et al., we evaluate style transfer across the three main properties of text style transfer: style transfer accuracy, content preservation, and fluency. We use our custom joint sequence classification models, implemented using HuggingFace libraries (Wolf et al., 2020) to evaluate the style transfer success ratio. Our definition for the Style Transfer Success  $S_{c}$  is the total number of matches between intended and transferred style buckets, divided by the total number of samples. To judge content preserved in style transferred text, we use three metrics: BLEU (Papineni et al., 2002), embedding-based similarity (Wieting et al., 2019) using cosine similarity of two sentence embeddings (Reimers and Gurevych, 2019), and Word Mover's Distance (WMD) (Mir et al., 2019). For fluency, we use measures like perplexity using GPT2 (Radford et al., 2019) and an adversarial classifier using the cross-aligned autoencoder model (Mir et al., 2019).",
+        "bbox": [
+            115,
+            192,
+            489,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Experimental Setup: In this paper, we illustrate different micro-style combinations in the training data, for a randomly selected case, with each combination in both the \"balanced\" and \"skewed settings. Therefore, we consider 6 cases respectively: 1) Formality and Arousal in a balanced setting (FA balanced) 2) Formality and Arousal in a skewed setting (FA skewed) 3) Formality, Arousal and Bias in a balanced setting (FAB balanced) 4) Formality, Arousal and Bias in skewed setting (FAB skewed) 5) Formality, Arousal, Bias and Sentiment in the balanced setting (FABS balanced) 6) Formality, Arousal, Bias and Sentiment in skewed setting (FABS skewed). We construct the training data with the appropriate settings and then pass them through our experimental pipeline (illustrated in Figure 2) and quantitatively evaluate the style transfer results.",
+        "bbox": [
+            115,
+            564,
+            489,
+            837
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Discussion: Table 2 shows examples of style-transferred sentences, given a style-transfer goal from our experimental pipeline for both balanced and skewed settings. E.g., given the objective is to decrease Formality but increase arousal, the sen",
+        "bbox": [
+            112,
+            838,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/1e9d114957a6fb214469fb38dffc5def353c5f7ec421f695d58efbfe4690f83e.jpg",
+        "image_caption": [
+            "Figure 3: Balancing micro-style distributions leads to a higher multi-style transfer percentage than in the Skewed setting in all the cases."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            80,
+            870,
+            189
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/34aaafb919848ec80a940e533a924f4347a57f86c667d381a9bd298128a21add.jpg",
+        "image_caption": [
+            "Figure 4: Considering the micro-style combinations such that, Formality [formal = f, informal = i], Bias [biased = b, unbiased = u], and Arousal [aroused = e, un-aroused = n], we observe that the micro-style combinations that are rarer (e.g., informal unbiased neutral (iun)) have more representation in the \"balanced\" setting than the \"skewed\" setting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            259,
+            870,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tence “Did you hear about the soldier with 8 limbs? He was army” transforms to “He’s an army soldier with 8 legs?”. Here, the contraction “He’s” indicates a formality decrease, and the replacement of limbs with legs indicates a decrease. The overall arousal of this sentence is higher when it transforms into a question.",
+        "bbox": [
+            505,
+            512,
+            882,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 illustrates that the balanced setup always has a higher success percentage of style transfer  $(S_{c})$  than the skewed setup. We cannot compare the success percentage across cases because matching the exact target and transferred style buckets becomes difficult as the number of micro-styles increases. We can also observe through Table 2 that the quality of the balanced transferred text aligns better with the style transfer goal than the skewed transferred text.",
+        "bbox": [
+            505,
+            627,
+            882,
+            788
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Figure 4, we compare the difference in representation percentage of specific style combinations in the test sample for a specific case where we consider Formality, Arousal, and Bias micro-styles. We observe that a balanced joint distribution leads to more representation in the style combinations that are less likely to occur. This is further accentuated as micro-styles increase, as reported in",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3935",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/c48ab5e0b9d5fd962040c10d371cd5beaaf8080af50b6d848531a4ca52b61713.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Style Transfer Goal</td><td>Input Text</td><td>Balanced Transferred Text</td><td>Skewed Transferred Text</td></tr><tr><td>↑ Formality ↑ Arousal</td><td>Wouldn’t it be great if Trump went 3rd party and sucked away millions of Republican votes lol</td><td>It wouldn’t be nice if Trump went to the third party and swooped millions of Repub-lican votes.</td><td>Would it not be great if Trump went 3rd party and sucked away millions of Republican votes?</td></tr><tr><td>↑ Formality ↓ Arousal</td><td>I didn’t know what happiness was until I got married. But by then it was too late.</td><td>Until I got married, I didn’t even know what happiness was.</td><td>I did not know what happiness was till I got married and it was too late.</td></tr><tr><td>↓ Formality ↑ Arousal</td><td>Did you hear about the soldier with 8 limbs? He was army</td><td>He’s an army soldier with 8 legs?</td><td>Did you hear about the soldier with 8 limbs in the army?</td></tr><tr><td>↓ Formality ↓ Arousal</td><td>Yeah, I don’t understand all the hate.</td><td>Yeah I’m not gonna understand the hate.</td><td>Yeah I do not understand all the hate.</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            885,
+            183
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Appendix C. In Figure 4, we see that rarer style combinations [ibn, fun, iun] show more representation in the balanced case as compared to the skewed case. This supports our intuition that the style transfer model benefits from learning the representation of all possible style combinations that can occur together.",
+        "bbox": [
+            110,
+            246,
+            487,
+            357
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "When we consider Formality, Arousal, and Bias micro styles together, the most represented category (30% of samples) is \"formal unbiased aroused\" (fue). The least represented category (as unlikely to occur together) is \"informal unbiased unaroused\" (iun) with 1%. We observe that the quantitative evaluation metrics are quite indicative when compared across style combinations. For instance, in Table 3, we observe that perplexity increases in categories that are unlikely to occur together (iun). This indicates that the style transfer model is confused by the style distributions present for this style combination.",
+        "bbox": [
+            110,
+            359,
+            489,
+            567
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We do not claim that our method of balancing multiple styles will work even for entangled micro-style combinations, as that is out of the scope of the current paper. However, balancing considerably affects the multi-style transfer output for the range of micro-style combinations we considered, and that has an application in many NLP tasks. This result could hugely influence future studies exploring better ways to balance even the entangled micro-styles.",
+        "bbox": [
+            112,
+            569,
+            489,
+            731
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            746,
+            247,
+            762
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multi-style text style transfer is a challenging problem predominantly plagued by the need for jointly annotated high-quality datasets. There is a clear need for more research about the marginal and joint distribution of inherent micro-styles present in the training dataset used for style transfer. Multi-style text-style transfer typically requires access to large, jointly labeled datasets and many computational resources under typical implementations. More",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e03ae57b2e65e4ddf26286d83ccac12ad2dfc11ca76829c89dc2406a648fbf8a.jpg",
+        "table_caption": [
+            "Table 2: The table shows the style transferred sentences, given an input sentence and the intended style transfer goal, for both the balanced setting as well as the skewed setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Setting</td><td>Styles</td><td>Perp</td><td>Adv</td><td>BLEU</td><td>Cos</td><td>WMD</td></tr><tr><td rowspan=\"2\">Balanced</td><td>fue</td><td>115.16</td><td>0.90</td><td>0.77</td><td>0.92</td><td>0.32</td></tr><tr><td>iun*</td><td>598.58</td><td>0.86</td><td>0.78</td><td>0.92</td><td>0.36</td></tr><tr><td rowspan=\"2\">Skewed</td><td>fue</td><td>116.02</td><td>0.90</td><td>0.78</td><td>0.92</td><td>0.32</td></tr><tr><td>iun*</td><td>650.47</td><td>0.82</td><td>0.77</td><td>0.93</td><td>0.37</td></tr></table>",
+        "bbox": [
+            515,
+            243,
+            878,
+            315
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Comparison of the evaluation metrics for the most represented style combination (fue - formal unbiased aroused) vs the least represented style combination (iun* - informal unbiased unaroused). One key observation is that perplexity increases when the style combinations are unlikely to occur together.",
+        "bbox": [
+            507,
+            324,
+            882,
+            411
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "importantly, we would not be able to conveniently tweak the input data distributions in other multi-style text style transfer methods.",
+        "bbox": [
+            507,
+            438,
+            882,
+            485
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we implement a multi-style transfer pipeline that subverts the requirement of a jointly annotated dataset of multiple styles by constructing a pseudo-parallel dataset to which we introduce our contribution of constructing style distributions. We then use the modified pseudo-parallel datasets for multi-style transfer. Our modified pipeline effectively allows us to understand the importance of the joint distribution of micro styles in training data and is a substantial contribution.",
+        "bbox": [
+            507,
+            487,
+            882,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We quantitatively explore the impact of joint micro-style distributions in the training dataset on the style-transferred output sentences. When the joint micro-style distributions are balanced, there is more control over style-transferred output than with a skewed distribution. These findings will likely inform the design of multi-style transfer datasets and encourage us to explore the micro-style relationships in our datasets.",
+        "bbox": [
+            507,
+            649,
+            884,
+            794
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            810,
+            613,
+            824
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this research, though we employed automatic evaluation of our multi-style transferred text, we acknowledge that multi-style transfer is challenging to observe with the existing metrics for style transfer evaluation, and human evaluation should",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3936",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "be done as well. As this research paper focuses on exploring the impact of style distributions in the training data on style-transferred output rather than developing a superior multi-style text transfer model, we use quantitative evaluation in this iteration of our paper. We hope that the large sample size and the consistency of applied metrics make our automated approach a reasonable way of evaluating the style transfer output.",
+        "bbox": [
+            112,
+            84,
+            489,
+            228
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This iteration of our paper aims to achieve multi-style transfer across multiple micro styles taken into consideration together as our contribution would aid in constructing a training dataset for multiple micro-style style transfers. We did not explore another exciting question of how balancing multiple micro styles in the training dataset might influence individual style transfer, which could be a promising future direction for our study.",
+        "bbox": [
+            112,
+            230,
+            489,
+            374
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We acknowledge that the classifier's quality sets an upper bound on the best style transfer accuracy that is obtainable. However, the target task is quite complicated without a parallel dataset. Our objective was not to have the most accurate classification of micro styles but to find a means to get acceptable pseudo labels for the micro styles. Individually, all our micro style classifiers had a classification accuracy of  $80\\%$  F1 and higher, and we deemed this good enough for pseudo-label creation.",
+        "bbox": [
+            112,
+            375,
+            489,
+            535
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also focused on utilizing the present styles in the training data and classifying them to derive inherent training style distributions instead of dynamically tuning the proportion of styles present in the training dataset. However, tuning these style proportions using techniques such as PPLM (Dathathri et al., 2019) would give us greater control over our experimental pipeline and is an appropriate next step.",
+        "bbox": [
+            112,
+            536,
+            489,
+            681
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            694,
+            278,
+            709
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We thank Vivek Aithal, Priyam Srivastava and Daniel McAndrew for their initial work on the pipeline for multi-style transfer. This was instrumental to our project and helped us get a kickstart on our research.",
+        "bbox": [
+            112,
+            720,
+            489,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            828,
+            213,
+            843
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "ref_text",
+        "text": "Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587-604.",
+        "bbox": [
+            114,
+            851,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sven Buechel and Udo Hahn. 2022. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. arXiv preprint arXiv:2205.01996.",
+            "Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2019. Plug and play language models: A simple approach to controlled text generation. arXiv preprint arXiv:1912.02164.",
+            "Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Exploration and evaluation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.",
+            "Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford, 1(12):2009.",
+            "Navita Goyal, Balaji Vasan Srinivasan, Anandhavelu Natarajan, and Abhilasha Sancheti. 2020. Multi-style transfer with discriminative feedback on disjoint corpus. arXiv preprint arXiv:2010.11578.",
+            "Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324.",
+            "Junxian He, Xinyi Wang, Graham Neubig, and Taylor Berg-Kirkpatrick. 2020. A probabilistic formulation of unsupervised text style transfer. arXiv preprint arXiv:2002.03912.",
+            "Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751.",
+            "Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In *The 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*. Association for Computational Linguistics.",
+            "Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In International conference on machine learning, pages 1587-1596. PMLR.",
+            "Harsh Jhamtani, Varun Gangal, Eduard Hovy, and Eric Nyberg. 2017. Shakespearizing modern language using copy-enriched sequence-to-sequence models. arXiv preprint arXiv:1707.01161.",
+            "Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea. 2022. Deep learning for text style transfer: A survey. Computational Linguistics, 48(1):155-205.",
+            "Dongyeop Kang and Eduard Hovy. 2019. xslue: A benchmark and analysis platform for cross-style language understanding and evaluation. arXiv preprint arXiv:1911.03663."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3937",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Dongyeop Kang and Eduard Hovy. 2021. Style is not a single variable: Case studies for cross-stylistic language understanding. 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 2376-2387.",
+            "Kalpesh Krishna, John Wieting, and Mohit Iyyer. 2020. Reformulating unsupervised style transfer as paraphrase generation. arXiv preprint arXiv:2010.05700.",
+            "Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y-Lan Boureau. 2018. Multiple-attribute text rewriting. In International Conference on Learning Representations.",
+            "Dianqi Li, Yizhe Zhang, Zhe Gan, Yu Cheng, Chris Brockett, Ming-Ting Sun, and Bill Dolan. 2019. Domain adaptive text style transfer. arXiv preprint arXiv:1908.09395.",
+            "Remi Mir, Bjarke Felbo, Nick Obradovich, and Iyad Rahwan. 2019. Evaluating style transfer for text. CoRR, abs/1904.02295.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.",
+            "Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, and Diyi Yang. 2020. Automatically neutralizing subjective bias in text. In Proceedings of the aaai conference on artificial intelligence, volume 34, pages 480-489.",
+            "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.",
+            "Sudha Rao and Joel Tetreault. 2018. Dear sir or madam, may i introduce the gyafc dataset: Corpus, benchmarks and metrics for formality style transfer. arXiv preprint arXiv:1803.06535.",
+            "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks.",
+            "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Stephanie Schoch, Wanyu Du, and Yangfeng Ji. 2021. Contextualizing variation in text style transfer datasets. arXiv preprint arXiv:2108.07871.",
+            "Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. Advances in neural information processing systems, 30.",
+            "Sandeep Subramanian, Guillaume Lample, Eric Michael Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y-Lan Boureau. 2018. Multiple-attribute text style transfer. arXiv preprint arXiv:1811.00552.",
+            "Enrica Troiano, Aswathy Velutharambath, et al. 2021. From theories on styles to their transfer in text: Bridging the gap with a hierarchical survey. arXiv preprint arXiv:2110.15871.",
+            "Ashwin K Vijayakumar, Michael Cogswell, Ramprasath R Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424.",
+            "Amy Beth Warriner, Victor Kuperman, and Marc Brysbaert. 2013. Norms of valence, arousal, and dominance for 13,915 english lemmas. Behavior research methods, 45(4):1191-1207.",
+            "Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural text generation with unlikelihood training. arXiv preprint arXiv:1908.04319.",
+            "John Wieting, Taylor Berg-Kirkpatrick, Kevin Gimpel, and Graham Neubig. 2019. Beyond bleu: training neural machine translation with semantic similarity. arXiv preprint arXiv:1909.06694.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45.",
+            "Yi Zhang, Tao Ge, and Xu Sun. 2020. Parallel data augmentation for formality style transfer. arXiv preprint arXiv:2005.07522."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            737
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Dataset Information",
+        "text_level": 1,
+        "bbox": [
+            510,
+            749,
+            722,
+            764
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We choose four micro-styles from both intended and unintended style categories, based on the style hierarchy as defined in Troiano et al. - Formality, Arousal, Sentiment, and Bias. While formality is considered a \"non-targeted intended\" micro-style, arousal and sentiment are \"targeted intended\" micro-styles. We also include subjective bias, an \"unintended\" micro-style, to ensure we include styles from all hierarchy branches. We built",
+        "bbox": [
+            509,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3938",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "our micro-style joint classification and style transfer models from multiple publicly available NLP datasets built by other researchers, and we detail these below.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Formality. We use Grammarly's Yahoo Answers Formality Corpus (Rao and Tetreault, 2018), which consists of  $105\\mathrm{k}$  sentences from two styles: \"formal\" and \"informal\" sentences written either in formal or informal modern English. Unlike formal sentences, informal sentences tend to have more misspellings, short forms (\"u\" instead of \"you\"), and non-standard usage of punctuation.",
+        "bbox": [
+            110,
+            168,
+            489,
+            297
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Arousal. We use the emotion annotated Emobank dataset (Buechel and Hahn, 2022) based on the three-dimensional VAD model developed by (Warriner et al., 2013). In particular, we transform the Arousal dimension into binary categories such as \"arousal\" and \"non-arousal.\"",
+        "bbox": [
+            112,
+            317,
+            487,
+            413
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Sentiment. We use the famous Sentiment140 dataset (Go et al., 2009), which consists of automatically annotated tweets, where the tweets containing positive emoticons are assumed as positive. In contrast, those with negative emoticons are assumed to be negative. The training dataset consisted of 1.6M tweets, and the test dataset consisted of 359 tweets. The tweets were preprocessed using NLTK to remove special Twitter-specific symbols like hashtags, usernames, and URLs.",
+        "bbox": [
+            112,
+            432,
+            489,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Bias. We use the Wiki Neutrality Corpus by (Pryzant et al., 2020). This is a new parallel corpus of 180,000 biased and neutralized sentence pairs. In order to train our joint classifier models, we used the training dataset from the appropriate micro-style datasets mentioned above. To implement our style distribution hypothesis, we used random samples for training and testing, from the combination of all the dev datasets from the benchmarks corpus by (Kang and Hovy, 2019). This consists of 15 different styles coupled to both content and domain by varying degrees. We wanted to ensure that the dataset used for training our style transfer model and verifying our hypothesis has sufficient indicators of the appropriate micro-styles. This could be done best by using a sample consisting of datasets curated for each individual micro-style (since a jointly annotated dataset with so many styles is not available).",
+        "bbox": [
+            112,
+            613,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B Multi Style Transfer Pipeline",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            798,
+            101
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.1 Resources used for Training",
+        "text_level": 1,
+        "bbox": [
+            509,
+            122,
+            781,
+            139
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "All models were trained using cloud GPUs on Google Colab Pro and Pro+. We used 1 V100 GPU in its \"High-RAM\" (52GB) GPU run-time setting to train the paraphrase generation model, while for other models we used 1 P100 GPU at the \"standard RAM\" setting (32GB).",
+        "bbox": [
+            507,
+            154,
+            884,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.2 Diverse Paraphrase Generation",
+        "text_level": 1,
+        "bbox": [
+            509,
+            280,
+            806,
+            294
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We adopted a pre-trained T5 model (Raffel et al., 2020), to generate paraphrases. We trained the model on the ParaNMT-filtered dataset provided by (Krishna et al., 2020). This is a subset of the ParaNMT dataset with filters applied to promote lexical diversity, syntactic diversity, and semantic similarity. This model was then used to generate the pseudo-parallel training data for transfer. We selected the t5-small architecture (60 million parameters) as this is approximately 10x smaller than the GPT-2 large model used in (Krishna et al., 2020). We used the hyper-parameters given in Table 4. Based on the recommendation in the appendix of Raffel et al, we used the \"paraphrase:\" prefix to train the paraphrase model. Once we had this trained paraphrase model, we used diverse beam search (Vijayakumar et al., 2016) to generate diverse paraphrased outputs. The hyper-parameters used for diverse beam search are mentioned in Table 5. We preferred beam search over top-p sampling in order to prioritize fluent paraphrases (Welleck et al., 2019) over unique paraphrases.",
+        "bbox": [
+            507,
+            311,
+            884,
+            667
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Input - paraphrase: I love to play my guitar and I do not know why",
+        "bbox": [
+            507,
+            687,
+            882,
+            719
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Output - I love playing my guitar and I'm not sure why",
+        "bbox": [
+            507,
+            719,
+            882,
+            751
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/457e58d4c10105bed3065841e9366d945fc9db54ce56da44dfad2c2d5b74ccea.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>batch size</td><td>8</td></tr><tr><td>number of epochs</td><td>12</td></tr><tr><td>learning rate</td><td>1e-4</td></tr><tr><td>max sequence length</td><td>64</td></tr></table>",
+        "bbox": [
+            531,
+            774,
+            863,
+            876
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Hyper parameters for T5 training for paraphrase generation",
+        "bbox": [
+            507,
+            885,
+            884,
+            914
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3939",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/177f9cc965e5eeb3eaba89e0f9592e09cf414ceb98af591b4a007a89e9380008.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>max length</td><td>70</td></tr><tr><td>early stopping</td><td>True</td></tr><tr><td>no repeat ngram size</td><td>5</td></tr><tr><td>num beams</td><td>9</td></tr><tr><td>num beam groups</td><td>3</td></tr><tr><td>diversity penalty</td><td>0.5</td></tr></table>",
+        "bbox": [
+            136,
+            80,
+            467,
+            214
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B.3 Micro-style Classification",
+        "text_level": 1,
+        "bbox": [
+            112,
+            256,
+            363,
+            273
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We trained a joint sentence classification model to classify the sentence on multiple axes inspired by the approach in Kang and Hovy, which uses an encoder-decoder-based model that learns cross-style patterns with the shared internal representation across styles. Our joint model comprises fully connected layers attached to a DistilBERT model (Sanh et al., 2019), which acts as an encoder. The hyperparameters for this joint model are given in Table 6. This single model effectively replaces the need for a different model for each classification task, significantly reducing the need for computing resources for training and inference. Our joint classifier is essential for downstream tasks like training style transfer models and evaluation. Say we first perform a joint classification of both formality and arousal micro-styles on our datasets, considering we want a multiple-style transfer along the axes of formality and arousal. This results in both formality and arousal pseudo-labels for the sentences. Since these labels are generated algorithmically rather than by hand, we refer to them as pseudo-labels. Pseudo-labeled sentences can then be used to generate the pseudo-parallel dataset for training joint style transfer models and directly measure the variation of a style along the axis of interest.",
+        "bbox": [
+            112,
+            278,
+            489,
+            713
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/84e6a5473622553be525963f7214c012dd73171272645d4a7685538ae29d2169.jpg",
+        "table_caption": [
+            "Table 5: Hyperparameters for Beam Search"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>train batch size</td><td>256</td></tr><tr><td>test batch size</td><td>512</td></tr><tr><td>number of epochs</td><td>3</td></tr><tr><td>learning rate</td><td>1e-4</td></tr></table>",
+        "bbox": [
+            136,
+            738,
+            467,
+            822
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B.4 Pseudo Parallel Data Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            866,
+            421,
+            881
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We then selected the best paraphrase (most stylistically different from the anchor sentence) based",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "on the cosine distance between the anchor and the paraphrased sentence's style vectors. To enable the transfer model to transfer to specified levels of a particular style, we defined 'very low', 'low', 'mid', 'high', and 'very high' buckets for each micro-style. In the following, we describe the bucket boundaries for our style scores.",
+        "bbox": [
+            507,
+            84,
+            884,
+            195
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Buckets: Very Low = [0, 0.2] Low = [0.2, 0.4]  \nMid = [0.4, 0.6] High = [0.6, 0.95] Very High = [0.95,1]",
+        "bbox": [
+            507,
+            197,
+            882,
+            244
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Using the absolute difference between original text style scores and their best-paraphrased text style scores, we find paraphrasing successfully stripped away both formality and arousal aspects of the text. The same phenomenon has been observed in previous studies, such as (Krishna et al., 2020). To ensure a diverse pseudo-parallel dataset, we retain only anchor-paraphrase pairs that do not match in terms of their style bucket. For example, if an anchor-paraphrase sentence pair is assigned style buckets for formality and arousal, as [very high, low] and [very high, very low], this pair will be retained. However, if both style buckets match, the sentence pair will not be considered diverse enough to remain in the pseudo-parallel dataset. In style transfer models, the paraphrased sentence and its style buckets are used as input parameters, while the style buckets assigned to the anchor sentence are used as proxy levels of output style transfer. The following is an example of the input to the joint style transfer model and the expected output.",
+        "bbox": [
+            507,
+            246,
+            884,
+            600
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Goal - Highly increase the formality of the sentence, slightly increase the arousal of the sentence  \nInput - transfer: I'm sad you're going | input formality: low | input arousal: low | output formality: high | output arousal: mid",
+        "bbox": [
+            507,
+            615,
+            884,
+            695
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Output - I am sorry you are going to go.",
+        "bbox": [
+            507,
+            697,
+            811,
+            713
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/8ae93d21788a5678f8d3268bcba99794e8f25c858a9bed15c636c5a21c5f1299.jpg",
+        "table_caption": [
+            "Table 6: Hyperparameters for Joint Classifier"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>train batch size</td><td>8</td></tr><tr><td>test batch size</td><td>8</td></tr><tr><td>number of epochs</td><td>5</td></tr><tr><td>learning rate</td><td>1e-4</td></tr></table>",
+        "bbox": [
+            531,
+            725,
+            863,
+            810
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Hyperparameters for T5 for Style Transfer",
+        "bbox": [
+            519,
+            820,
+            870,
+            834
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B.5 Style Transfer Training",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            742,
+            881
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our T5 models were trained on pseudo-parallel datasets created and filtered as described earlier.",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3940",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/77c2fdd4393c5fcea492bfe16dcec1804e4b64e23d60a0c3a02891931353a5f8.jpg",
+        "image_caption": [
+            "Figure 5: Considering the micro-style combinations such that, Formality [formal = f, informal = i], Bias [biased = b, unbiased = u], Arousal [aroused = e, un-aroused = u], and Sentiment [negative = n, positive = p]; we observe that the micro-style combinations that are rarer have more representation in the \"balanced\" setting than the \"skewed\" setting. The categories fbnp, funp and iben have more representation for balanced setting vs skewed setting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            191,
+            82,
+            803,
+            275
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "According to the task, we converted the classifier predictions into buckets of style based on the chosen style of the original and then paraphrased sentences. To achieve the necessary intensity transfers, we appended this information to the paraphrased sentence, as motivated by the original T5 paper (Raffel et al., 2020). Hyperparameters are mentioned in Table 7. For generating sentences from our trained models, we used a combination of both beam search (Vijayakumar et al., 2016) and nucleus sampling (Holtzman et al., 2019) and chose the top 3 sentences from the generations.",
+        "bbox": [
+            112,
+            370,
+            489,
+            565
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "C Some Additional Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            577,
+            366,
+            593
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "C.1 Impact of Fluency filter on input training data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            605,
+            485,
+            634
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We find that filtering the original dataset based on fluency metrics always results in better style transferred output as compared to the transferred output when the input dataset is not filtered. This is intuitive, as better quality input prevents confusion in the style transfer model and leads to better quality output. As a result of this finding, we use a fluency filter (adversarial classifier  $>0.1$  and perplexity  $< 365$ ), before we conduct any of the rest of our experiments with micro-style distributions.",
+        "bbox": [
+            112,
+            643,
+            489,
+            803
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "C.2 Balancing effect on lesser represented style combinations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            816,
+            460,
+            848
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "In Figure 6, we consider the case where we examine Formality [formal = f, informal = i] and Arousal [aroused = e, un-aroused = u] micro-styles and compare the percentage of specific style combinations",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/924bd3451ecd9865043c114d2e82e26722acaff2ce9606d5513b4898241e5448.jpg",
+        "image_caption": [
+            "Figure 6: Considering the micro-style combinations such that, Formality [formal = f, informal = i] and Arousal [aroused = e, un-aroused = u]; we observe that the micro-style combinations that are rare (ie, in) have more representation in the \"balanced\" setting than the \"skewed\" setting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            368,
+            870,
+            479
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "in the test sample. We observe that as the number of micro styles increases, a balanced joint distribution leads to more representation in combinations that are less likely to occur such as in or 'informal and neutral'.",
+        "bbox": [
+            507,
+            601,
+            882,
+            678
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Figure 5 shows a similarly pronounced effect. Here the number of micro styles is increased, and we can observe that the balanced setting shows higher representation than the skewed setting. An example of an unlikely style combination is fbnp, or \"formal biased neutral and positive\". We also observe that as the number of micro-style increases, there is no representation in some combinations in both settings [ibnp, iunn, iuen, iunp]. This is a natural result as some micro-style combinations cannot exist in nature.",
+        "bbox": [
+            507,
+            681,
+            884,
+            857
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3941",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? 5",
+            "A2. Did you discuss any potential risks of your work? 5",
+            "A3. Do the abstract and introduction summarize the paper's main claims?",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            288
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            298,
+            489,
+            316
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "2,3",
+        "bbox": [
+            132,
+            321,
+            159,
+            334
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? 2,3",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        ],
+        "bbox": [
+            127,
+            346,
+            880,
+            752
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "3,B, C",
+        "bbox": [
+            132,
+            785,
+            184,
+            800
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? B.1",
+        "bbox": [
+            129,
+            810,
+            880,
+            859
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            866,
+            877,
+            889
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3942",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            129,
+            83,
+            878,
+            115
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B",
+        "bbox": [
+            151,
+            117,
+            166,
+            129
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            129,
+            142,
+            880,
+            190
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B",
+        "bbox": [
+            151,
+            192,
+            166,
+            204
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            129,
+            217,
+            880,
+            265
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A",
+        "bbox": [
+            149,
+            268,
+            166,
+            280
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            340,
+            880,
+            372
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            373,
+            349,
+            388
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            399,
+            880,
+            447
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            448,
+            349,
+            463
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            475,
+            880,
+            521
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            523,
+            349,
+            539
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            549,
+            873,
+            564
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            565,
+            349,
+            581
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            592,
+            880,
+            623
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            625,
+            349,
+            640
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "3943",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 11
+    }
+]

2023/Balancing the Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer/e52f8ce0-4325-4d7a-9b9b-68ca012cb308_model.json

@@ -0,0 +1,2259 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.134,
+                0.09,
+                0.868,
+                0.131
+            ],
+            "angle": 0,
+            "content": "Balancing Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.306,
+                0.156,
+                0.699,
+                0.173
+            ],
+            "angle": 0,
+            "content": "Debarati Das David Ma Dongyeop Kang"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.262,
+                0.174,
+                0.74,
+                0.189
+            ],
+            "angle": 0,
+            "content": "Department of Computer Science, University of Minnesota"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.306,
+                0.19,
+                0.698,
+                0.205
+            ],
+            "angle": 0,
+            "content": "{das00015, maxxx818, dongyeop}@umn.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.281,
+                0.461,
+                0.622
+            ],
+            "angle": 0,
+            "content": "Text style transfer is an exciting task within the field of natural language generation that is often plagued by the need for high-quality paired datasets. Furthermore, training a model for multi-attribute text style transfer requires datasets with sufficient support across all combinations of the considered stylistic attributes, adding to the challenges of training a style transfer model. This paper explores the impact of training data input diversity on the quality of the generated text from the multi-style transfer model. We construct a pseudo-parallel dataset by devising heuristics to adjust the style distribution in the training samples. We balance our training dataset using marginal and joint distributions to train our style transfer models. We observe that a balanced dataset produces more effective control effects over multiple styles than an imbalanced or skewed one. Through quantitative analysis, we explore the impact of multiple style distributions in training data on style-transferred output. These findings will better inform the design of style-transfer datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.636,
+                0.26,
+                0.652
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.662,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Multi-style text transfer is a challenging task today with applications such as automatic domain-appropriate, style-conformant writing (Fu et al., 2018) and AI-assisted stylistic language editing. Text style transfer is an intricate task as all language has a specific context, and those contexts influence the attributes of the language (Hovy and Yang, 2021). Text style transfer is challenging because it involves dealing with the aspects of style coupled with the textual content (Hu et al., 2017; Shen et al., 2017; Lample et al., 2018). This domain's other obstacles include the need for parallel corpus (Jhamtani et al., 2017) and quality training data. As the number of style dimensions increases with multi-style text transfer, not only is the requirement of a jointly annotated corpus across all the"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.537,
+                0.256,
+                0.885,
+                0.395
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.421,
+                0.886,
+                0.55
+            ],
+            "angle": 0,
+            "content": "Figure 1: When an input sentence is passed to the multi-style transfer model, to increase formality and decrease arousal, we hypothesize that when the model is trained on a balanced joint distribution of formality and arousal (all four style combinations have a \\(25\\%\\) representation) - the style transfer is more successful as opposed to when the model is trained on a skewed joint distribution (there is no representation of the \"informal unaroused\" style combination) of styles in the training data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.579,
+                0.882,
+                0.611
+            ],
+            "angle": 0,
+            "content": "stylistic dimensions problematic, but the different styles are not necessarily independent."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.612,
+                0.884,
+                0.869
+            ],
+            "angle": 0,
+            "content": "While \"style\" can also refer to authorial or domain-specific style, in this paper, we focus on \"microstyles\" as defined by (Kang and Hovy, 2021) where they define \"micro-style\" as a complex combination of different factors such as formality markers, emotions, and metaphors. People intentionally (Troiano et al., 2021) tune these styles in writing differently based on their mood, the person they are addressing, the content of the message, or the platform. Multiple microstyles can jointly describe a text; for example, a given text could simultaneously be formal and sad. Microstyles also more easily lend themselves to being represented as spectra with varying degrees of intensity. These points align with our vision of an application where users can edit micro-style aspects of their writing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.871,
+                0.884,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Much research exists on models implementing multi-style text transfer and interdependency of micro-styles (Kang and Hovy, 2019; Goyal et al.,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3932"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.229,
+                0.946,
+                0.769,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 3932-3943"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.297,
+                0.959,
+                0.7,
+                0.972
+            ],
+            "angle": 0,
+            "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.198,
+                0.086,
+                0.803,
+                0.304
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.316,
+                0.884,
+                0.361
+            ],
+            "angle": 0,
+            "content": "Figure 2: The input sentence transitions through every step in our multi-style text style transfer pipeline. The box in red indicates our main contribution to the pipeline, which helps us explore the effects of joint micro-style combinations on style-transferred output."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.384,
+                0.49,
+                0.658
+            ],
+            "angle": 0,
+            "content": "2020; Subramanian et al., 2018). However, there needs to be more exploration of the joint distribution of inherent micro-styles in the style transfer training dataset and how these micro-style distributions are related. Therefore, we pose a question - Can a dataset with minimal variance across multiple micro-style combinations, such that it experiences a \"balancing effect\", lead to a better style transferred output? Figure 1 illustrates our intuition that a dataset that experiences a \"balancing effect\" will have more control over the multi-style transferred output than a \"skewed\" dataset. Suppose the style transfer model sees examples of every style combination that can exist - this could aid in the style generation of even unlikely combinations of styles compared to a skewed distribution of these joint micro-styles."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.662,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this research, we consider a multi-style text style transfer pipeline assuming that the user has no access to parallel data or the style of the original text that he wishes to transfer, as would seem natural for a style language editing application. We introduce the changing of the training dataset micro-style joint distributions in such a pipeline and quantitatively explore the impact of this modification on the style transferred output. We perform a set of empirical analyses to demonstrate the influence of joint distributions on style-transferred output and show how this trend varies as the number of micro-styles considered changes. The 'balancing effect' on a training dataset leads to style transferred sentences from even the joint style combinations that are typically rare (\"informal unbiased and"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.384,
+                0.885,
+                0.545
+            ],
+            "angle": 0,
+            "content": "unaroused\"). Our study is the first of its kind on the distribution of micro styles in training datasets for multi-style text style transfer and is likely to have implications for designing datasets for multi-style transfer model training and fall within the context of and align with recent work on characterizing datasets and factors impacting style transfer (Bender and Friedman, 2018; Schoch et al., 2021; Li et al., 2019; Zhang et al., 2020; Gururangan et al., 2018)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.557,
+                0.795,
+                0.573
+            ],
+            "angle": 0,
+            "content": "2 Multi Style Transfer Pipeline"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.582,
+                0.884,
+                0.726
+            ],
+            "angle": 0,
+            "content": "Datasets: We chose four micro-styles from the style hierarchy defined in Troiano et al.: Formality, Arousal, Sentiment, and Bias, for our study and used publicly available NLP datasets built by other researchers (Rao and Tetreault, 2018; Buechel and Hahn, 2022; Go et al., 2009; Pryzant et al., 2020; Kang and Hovy, 2019) to develop and test our models. Appendix A mentions the details of the datasets and their usage."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.727,
+                0.885,
+                0.886
+            ],
+            "angle": 0,
+            "content": "Pipeline Overview: Our experimental setup for multi-style transfer is inspired by the work of (Krishna et al., 2020). Like them, we first generate a \"diverse\" paraphrase of the input sentence, and then the paraphrased sentence is rewritten in the style of choice. Towards this end, we train a paraphrasing model (separately on a parallel paraphrase dataset). Then, the trained paraphrase model is used to create \"pseudo-gold\" parallel data for training style models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "First, we adopted a pre-trained T5 model (Raffel et al., 2020) to generate paraphrases. This model"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3933"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.489,
+                0.213
+            ],
+            "angle": 0,
+            "content": "was trained for the task of paraphrase generation on the ParaNMT-filtered dataset provided by (Krishna et al., 2020). Once we had this trained paraphrase model, we used diverse beam search (Vijayakumar et al., 2016) to generate diverse fluent paraphrased outputs. An important assumption is that the paraphrase is stripped of its original style and does not leak into the training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.218,
+                0.49,
+                0.444
+            ],
+            "angle": 0,
+            "content": "We address this potential issue by training classifiers (Sanh et al., 2019) to predict style on the original and paraphrased datasets and find that all our micro-style classifiers have a classification accuracy of higher than \\(80\\%\\) F1, which is acceptable for pseudo-label creation. After we generate diverse paraphrases, we choose the most diverse paraphrase and then derive micro-style classifications for the paraphrased sentence using our trained micro-style classifiers. Therefore each sentence is assigned a classification score for each micro-style label and can form a \"pseudo parallel\" dataset for training the T5-based joint transfer model. Thus, our approach does not need a parallel dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.449,
+                0.49,
+                0.835
+            ],
+            "angle": 0,
+            "content": "We then converted the classifier predictions into buckets of style (ranging from \"very low\" to \"very high\") based on the chosen style of the original and then paraphrased sentences. The bucketing process is described in Appendix B. After this step, we introduce our contribution of \"constructing style distributions\" into the pipeline, as illustrated in Figure 2. Following that, we perform multi-style text style transfer. We appended the \"bucket\" information to the paraphrased sentence to achieve the necessary intensity transfers, as motivated by the original T5 paper (Raffel et al., 2020). We train T5-based style transfer models, where the paraphrased sentence and its style buckets are used as input parameters, while the style buckets assigned to the anchor sentence are used as proxy levels of output style transfer. All model-specific details are provided in Appendix B. For generating sentences from our trained models, we used beam search (Vijayakumar et al., 2016) and nucleus sampling (Holtzman et al., 2019) and chose the top 3 sentences from the generations. The following is an example of the input to the joint style transfer model and the expected output."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.488,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Goal - Highly increase the formality of the sentence, slightly increase the arousal of the sentence"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.904,
+                0.488,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Input - transfer: I'm sad you're going | input"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.883,
+                0.134
+            ],
+            "angle": 0,
+            "content": "formality: low | input arousal: low | output formality: high | output arousal: mid Output - I am sorry you are going to go."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.149,
+                0.885,
+                0.246
+            ],
+            "angle": 0,
+            "content": "Thus, we implemented a multi-style transfer pipeline to test our hypothesis without any finicky modeling paradigms popular in style transfer research, such as variational inference or autoregressive sampling (He et al., 2020; Subramanian et al., 2018)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.522,
+                0.258,
+                0.875,
+                0.351
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Style Combination</td><td>Balanced</td><td>Skewed</td></tr><tr><td>Formal Aroused</td><td>3395</td><td>8685</td></tr><tr><td>Formal Unaroused</td><td>3395</td><td>2792</td></tr><tr><td>Informal Aroused</td><td>3395</td><td>1275</td></tr><tr><td>Informal Unaroused</td><td>3395</td><td>828</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.36,
+                0.884,
+                0.403
+            ],
+            "angle": 0,
+            "content": "Table 1: Training data statistics (number of samples) for the balanced and skewed settings, when considering the micro-styles of Formality and Arousal."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.423,
+                0.885,
+                0.614
+            ],
+            "angle": 0,
+            "content": "Constructing Micro-style Distributions We define a \"style combination\" as a possible combination of the states that the micro-styles can take together - such as 'informal biased negative.' Since there are three micro-styles, each having binary states, the total possible number of style combinations, in this case, is given by \\( N_{c} = 2 \\times 2 \\times 2 = 2^{3} \\). Therefore to generalize, if \\( |m_{i}| \\) indicates the cardinality of each micro-style and \\( n \\) indicates the number of micro-styles considered, the total possible number of style combinations \\( (N_{c}) \\) possible is given by:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.641,
+                0.616,
+                0.884,
+                0.656
+            ],
+            "angle": 0,
+            "content": "\\[\nN _ {c} = \\prod_ {i = 1} ^ {n} | m _ {i} | \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.662,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "To create the balanced joint distribution of styles, we ensure the standard deviation across the style combinations is close to 0. We do this by down-sampling each style combination, such that the number of samples in each style combination is the same as the least represented style combination. As we increase micro-styles, some micro-style combinations do not occur naturally together, so their representation is close to 0. In such cases, we assume that the least represented style combination is at least \\(5\\%\\) of the total dataset. To ensure our comparison across the \"balanced\" and \"skew\" settings is fair, we construct a skewed dataset with a total sample size that is the same as that of the balanced dataset. Thus, the balanced dataset has a uniform distribution, while the skewed dataset"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3934"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.149
+            ],
+            "angle": 0,
+            "content": "has a non-uniform distribution. Table 1 shows the number of samples in each style combination of Formality and Arousal, given a \"balanced\" and \"skewed\" setting."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.165,
+                0.468,
+                0.182
+            ],
+            "angle": 0,
+            "content": "3 Experimental Results and Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.193,
+                0.49,
+                0.562
+            ],
+            "angle": 0,
+            "content": "Evaluation Metrics: Style transfer accuracy metrics quantify how nicely output texts match the desired style. However, more than this metric is required. Motivated by Jin et al., we evaluate style transfer across the three main properties of text style transfer: style transfer accuracy, content preservation, and fluency. We use our custom joint sequence classification models, implemented using HuggingFace libraries (Wolf et al., 2020) to evaluate the style transfer success ratio. Our definition for the Style Transfer Success \\( S_{c} \\) is the total number of matches between intended and transferred style buckets, divided by the total number of samples. To judge content preserved in style transferred text, we use three metrics: BLEU (Papineni et al., 2002), embedding-based similarity (Wieting et al., 2019) using cosine similarity of two sentence embeddings (Reimers and Gurevych, 2019), and Word Mover's Distance (WMD) (Mir et al., 2019). For fluency, we use measures like perplexity using GPT2 (Radford et al., 2019) and an adversarial classifier using the cross-aligned autoencoder model (Mir et al., 2019)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.565,
+                0.49,
+                0.838
+            ],
+            "angle": 0,
+            "content": "Experimental Setup: In this paper, we illustrate different micro-style combinations in the training data, for a randomly selected case, with each combination in both the \"balanced\" and \"skewed settings. Therefore, we consider 6 cases respectively: 1) Formality and Arousal in a balanced setting (FA balanced) 2) Formality and Arousal in a skewed setting (FA skewed) 3) Formality, Arousal and Bias in a balanced setting (FAB balanced) 4) Formality, Arousal and Bias in skewed setting (FAB skewed) 5) Formality, Arousal, Bias and Sentiment in the balanced setting (FABS balanced) 6) Formality, Arousal, Bias and Sentiment in skewed setting (FABS skewed). We construct the training data with the appropriate settings and then pass them through our experimental pipeline (illustrated in Figure 2) and quantitatively evaluate the style transfer results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.839,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Discussion: Table 2 shows examples of style-transferred sentences, given a style-transfer goal from our experimental pipeline for both balanced and skewed settings. E.g., given the objective is to decrease Formality but increase arousal, the sen"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.523,
+                0.082,
+                0.871,
+                0.19
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.2,
+                0.883,
+                0.243
+            ],
+            "angle": 0,
+            "content": "Figure 3: Balancing micro-style distributions leads to a higher multi-style transfer percentage than in the Skewed setting in all the cases."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.522,
+                0.26,
+                0.871,
+                0.372
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.383,
+                0.885,
+                0.483
+            ],
+            "angle": 0,
+            "content": "Figure 4: Considering the micro-style combinations such that, Formality [formal = f, informal = i], Bias [biased = b, unbiased = u], and Arousal [aroused = e, un-aroused = n], we observe that the micro-style combinations that are rarer (e.g., informal unbiased neutral (iun)) have more representation in the \"balanced\" setting than the \"skewed\" setting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.513,
+                0.884,
+                0.626
+            ],
+            "angle": 0,
+            "content": "tence “Did you hear about the soldier with 8 limbs? He was army” transforms to “He’s an army soldier with 8 legs?”. Here, the contraction “He’s” indicates a formality decrease, and the replacement of limbs with legs indicates a decrease. The overall arousal of this sentence is higher when it transforms into a question."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.628,
+                0.884,
+                0.789
+            ],
+            "angle": 0,
+            "content": "Figure 3 illustrates that the balanced setup always has a higher success percentage of style transfer \\((S_{c})\\) than the skewed setup. We cannot compare the success percentage across cases because matching the exact target and transferred style buckets becomes difficult as the number of micro-styles increases. We can also observe through Table 2 that the quality of the balanced transferred text aligns better with the style transfer goal than the skewed transferred text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In Figure 4, we compare the difference in representation percentage of specific style combinations in the test sample for a specific case where we consider Formality, Arousal, and Bias micro-styles. We observe that a balanced joint distribution leads to more representation in the style combinations that are less likely to occur. This is further accentuated as micro-styles increase, as reported in"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3935"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.115,
+                0.081,
+                0.887,
+                0.184
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Style Transfer Goal</td><td>Input Text</td><td>Balanced Transferred Text</td><td>Skewed Transferred Text</td></tr><tr><td>↑ Formality ↑ Arousal</td><td>Wouldn’t it be great if Trump went 3rd party and sucked away millions of Republican votes lol</td><td>It wouldn’t be nice if Trump went to the third party and swooped millions of Repub-lican votes.</td><td>Would it not be great if Trump went 3rd party and sucked away millions of Republican votes?</td></tr><tr><td>↑ Formality ↓ Arousal</td><td>I didn’t know what happiness was until I got married. But by then it was too late.</td><td>Until I got married, I didn’t even know what happiness was.</td><td>I did not know what happiness was till I got married and it was too late.</td></tr><tr><td>↓ Formality ↑ Arousal</td><td>Did you hear about the soldier with 8 limbs? He was army</td><td>He’s an army soldier with 8 legs?</td><td>Did you hear about the soldier with 8 limbs in the army?</td></tr><tr><td>↓ Formality ↓ Arousal</td><td>Yeah, I don’t understand all the hate.</td><td>Yeah I’m not gonna understand the hate.</td><td>Yeah I do not understand all the hate.</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.192,
+                0.884,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Table 2: The table shows the style transferred sentences, given an input sentence and the intended style transfer goal, for both the balanced setting as well as the skewed setting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.247,
+                0.489,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Appendix C. In Figure 4, we see that rarer style combinations [ibn, fun, iun] show more representation in the balanced case as compared to the skewed case. This supports our intuition that the style transfer model benefits from learning the representation of all possible style combinations that can occur together."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.36,
+                0.49,
+                0.568
+            ],
+            "angle": 0,
+            "content": "When we consider Formality, Arousal, and Bias micro styles together, the most represented category (30% of samples) is \"formal unbiased aroused\" (fue). The least represented category (as unlikely to occur together) is \"informal unbiased unaroused\" (iun) with 1%. We observe that the quantitative evaluation metrics are quite indicative when compared across style combinations. For instance, in Table 3, we observe that perplexity increases in categories that are unlikely to occur together (iun). This indicates that the style transfer model is confused by the style distributions present for this style combination."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.57,
+                0.49,
+                0.732
+            ],
+            "angle": 0,
+            "content": "We do not claim that our method of balancing multiple styles will work even for entangled micro-style combinations, as that is out of the scope of the current paper. However, balancing considerably affects the multi-style transfer output for the range of micro-style combinations we considered, and that has an application in many NLP tasks. This result could hugely influence future studies exploring better ways to balance even the entangled micro-styles."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.747,
+                0.248,
+                0.763
+            ],
+            "angle": 0,
+            "content": "4 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Multi-style text style transfer is a challenging problem predominantly plagued by the need for jointly annotated high-quality datasets. There is a clear need for more research about the marginal and joint distribution of inherent micro-styles present in the training dataset used for style transfer. Multi-style text-style transfer typically requires access to large, jointly labeled datasets and many computational resources under typical implementations. More"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.516,
+                0.244,
+                0.88,
+                0.316
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Setting</td><td>Styles</td><td>Perp</td><td>Adv</td><td>BLEU</td><td>Cos</td><td>WMD</td></tr><tr><td rowspan=\"2\">Balanced</td><td>fue</td><td>115.16</td><td>0.90</td><td>0.77</td><td>0.92</td><td>0.32</td></tr><tr><td>iun*</td><td>598.58</td><td>0.86</td><td>0.78</td><td>0.92</td><td>0.36</td></tr><tr><td rowspan=\"2\">Skewed</td><td>fue</td><td>116.02</td><td>0.90</td><td>0.78</td><td>0.92</td><td>0.32</td></tr><tr><td>iun*</td><td>650.47</td><td>0.82</td><td>0.77</td><td>0.93</td><td>0.37</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.325,
+                0.884,
+                0.412
+            ],
+            "angle": 0,
+            "content": "Table 3: Comparison of the evaluation metrics for the most represented style combination (fue - formal unbiased aroused) vs the least represented style combination (iun* - informal unbiased unaroused). One key observation is that perplexity increases when the style combinations are unlikely to occur together."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.439,
+                0.884,
+                0.486
+            ],
+            "angle": 0,
+            "content": "importantly, we would not be able to conveniently tweak the input data distributions in other multi-style text style transfer methods."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.488,
+                0.884,
+                0.648
+            ],
+            "angle": 0,
+            "content": "In this paper, we implement a multi-style transfer pipeline that subverts the requirement of a jointly annotated dataset of multiple styles by constructing a pseudo-parallel dataset to which we introduce our contribution of constructing style distributions. We then use the modified pseudo-parallel datasets for multi-style transfer. Our modified pipeline effectively allows us to understand the importance of the joint distribution of micro styles in training data and is a substantial contribution."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.65,
+                0.885,
+                0.795
+            ],
+            "angle": 0,
+            "content": "We quantitatively explore the impact of joint micro-style distributions in the training dataset on the style-transferred output sentences. When the joint micro-style distributions are balanced, there is more control over style-transferred output than with a skewed distribution. These findings will likely inform the design of multi-style transfer datasets and encourage us to explore the micro-style relationships in our datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.811,
+                0.615,
+                0.825
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this research, though we employed automatic evaluation of our multi-style transferred text, we acknowledge that multi-style transfer is challenging to observe with the existing metrics for style transfer evaluation, and human evaluation should"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3936"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.229
+            ],
+            "angle": 0,
+            "content": "be done as well. As this research paper focuses on exploring the impact of style distributions in the training data on style-transferred output rather than developing a superior multi-style text transfer model, we use quantitative evaluation in this iteration of our paper. We hope that the large sample size and the consistency of applied metrics make our automated approach a reasonable way of evaluating the style transfer output."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.231,
+                0.49,
+                0.375
+            ],
+            "angle": 0,
+            "content": "This iteration of our paper aims to achieve multi-style transfer across multiple micro styles taken into consideration together as our contribution would aid in constructing a training dataset for multiple micro-style style transfers. We did not explore another exciting question of how balancing multiple micro styles in the training dataset might influence individual style transfer, which could be a promising future direction for our study."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.376,
+                0.49,
+                0.536
+            ],
+            "angle": 0,
+            "content": "We acknowledge that the classifier's quality sets an upper bound on the best style transfer accuracy that is obtainable. However, the target task is quite complicated without a parallel dataset. Our objective was not to have the most accurate classification of micro styles but to find a means to get acceptable pseudo labels for the micro styles. Individually, all our micro style classifiers had a classification accuracy of \\(80\\%\\) F1 and higher, and we deemed this good enough for pseudo-label creation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.537,
+                0.49,
+                0.682
+            ],
+            "angle": 0,
+            "content": "We also focused on utilizing the present styles in the training data and classifying them to derive inherent training style distributions instead of dynamically tuning the proportion of styles present in the training dataset. However, tuning these style proportions using techniques such as PPLM (Dathathri et al., 2019) would give us greater control over our experimental pipeline and is an appropriate next step."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.695,
+                0.279,
+                0.71
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.721,
+                0.49,
+                0.801
+            ],
+            "angle": 0,
+            "content": "We thank Vivek Aithal, Priyam Srivastava and Daniel McAndrew for their initial work on the pipeline for multi-style transfer. This was instrumental to our project and helped us get a kickstart on our research."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.829,
+                0.214,
+                0.844
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.115,
+                0.852,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587-604."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.14
+            ],
+            "angle": 0,
+            "content": "Sven Buechel and Udo Hahn. 2022. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. arXiv preprint arXiv:2205.01996."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.15,
+                0.885,
+                0.216
+            ],
+            "angle": 0,
+            "content": "Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2019. Plug and play language models: A simple approach to controlled text generation. arXiv preprint arXiv:1912.02164."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.227,
+                0.885,
+                0.28
+            ],
+            "angle": 0,
+            "content": "Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Exploration and evaluation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.291,
+                0.885,
+                0.331
+            ],
+            "angle": 0,
+            "content": "Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford, 1(12):2009."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.342,
+                0.885,
+                0.395
+            ],
+            "angle": 0,
+            "content": "Navita Goyal, Balaji Vasan Srinivasan, Anandhavelu Natarajan, and Abhilasha Sancheti. 2020. Multi-style transfer with discriminative feedback on disjoint corpus. arXiv preprint arXiv:2010.11578."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.405,
+                0.885,
+                0.459
+            ],
+            "angle": 0,
+            "content": "Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.47,
+                0.885,
+                0.522
+            ],
+            "angle": 0,
+            "content": "Junxian He, Xinyi Wang, Graham Neubig, and Taylor Berg-Kirkpatrick. 2020. A probabilistic formulation of unsupervised text style transfer. arXiv preprint arXiv:2002.03912."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.533,
+                0.885,
+                0.574
+            ],
+            "angle": 0,
+            "content": "Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.584,
+                0.885,
+                0.663
+            ],
+            "angle": 0,
+            "content": "Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In *The 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.674,
+                0.885,
+                0.728
+            ],
+            "angle": 0,
+            "content": "Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In International conference on machine learning, pages 1587-1596. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.738,
+                0.885,
+                0.791
+            ],
+            "angle": 0,
+            "content": "Harsh Jhamtani, Varun Gangal, Eduard Hovy, and Eric Nyberg. 2017. Shakespearizing modern language using copy-enriched sequence-to-sequence models. arXiv preprint arXiv:1707.01161."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.802,
+                0.885,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea. 2022. Deep learning for text style transfer: A survey. Computational Linguistics, 48(1):155-205."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.866,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Dongyeop Kang and Eduard Hovy. 2019. xslue: A benchmark and analysis platform for cross-style language understanding and evaluation. arXiv preprint arXiv:1911.03663."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3937"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.179
+            ],
+            "angle": 0,
+            "content": "Dongyeop Kang and Eduard Hovy. 2021. Style is not a single variable: Case studies for cross-stylistic language understanding. 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 2376-2387."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.19,
+                0.49,
+                0.231
+            ],
+            "angle": 0,
+            "content": "Kalpesh Krishna, John Wieting, and Mohit Iyyer. 2020. Reformulating unsupervised style transfer as paraphrase generation. arXiv preprint arXiv:2010.05700."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.242,
+                0.488,
+                0.307
+            ],
+            "angle": 0,
+            "content": "Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y-Lan Boureau. 2018. Multiple-attribute text rewriting. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.32,
+                0.488,
+                0.372
+            ],
+            "angle": 0,
+            "content": "Dianqi Li, Yizhe Zhang, Zhe Gan, Yu Cheng, Chris Brockett, Ming-Ting Sun, and Bill Dolan. 2019. Domain adaptive text style transfer. arXiv preprint arXiv:1908.09395."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.385,
+                0.488,
+                0.424
+            ],
+            "angle": 0,
+            "content": "Remi Mir, Bjarke Felbo, Nick Obradovich, and Iyad Rahwan. 2019. Evaluating style transfer for text. CoRR, abs/1904.02295."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.437,
+                0.488,
+                0.529
+            ],
+            "angle": 0,
+            "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.541,
+                0.488,
+                0.607
+            ],
+            "angle": 0,
+            "content": "Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, and Diyi Yang. 2020. Automatically neutralizing subjective bias in text. In Proceedings of the aaai conference on artificial intelligence, volume 34, pages 480-489."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.619,
+                0.488,
+                0.672
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.684,
+                0.488,
+                0.749
+            ],
+            "angle": 0,
+            "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.762,
+                0.488,
+                0.815
+            ],
+            "angle": 0,
+            "content": "Sudha Rao and Joel Tetreault. 2018. Dear sir or madam, may i introduce the gyafc dataset: Corpus, benchmarks and metrics for formality style transfer. arXiv preprint arXiv:1803.06535."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.827,
+                0.488,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Stephanie Schoch, Wanyu Du, and Yangfeng Ji. 2021. Contextualizing variation in text style transfer datasets. arXiv preprint arXiv:2108.07871."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.135,
+                0.883,
+                0.189
+            ],
+            "angle": 0,
+            "content": "Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. Advances in neural information processing systems, 30."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.198,
+                0.883,
+                0.263
+            ],
+            "angle": 0,
+            "content": "Sandeep Subramanian, Guillaume Lample, Eric Michael Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y-Lan Boureau. 2018. Multiple-attribute text style transfer. arXiv preprint arXiv:1811.00552."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.273,
+                0.883,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Enrica Troiano, Aswathy Velutharambath, et al. 2021. From theories on styles to their transfer in text: Bridging the gap with a hierarchical survey. arXiv preprint arXiv:2110.15871."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.335,
+                0.883,
+                0.401
+            ],
+            "angle": 0,
+            "content": "Ashwin K Vijayakumar, Michael Cogswell, Ramprasath R Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.41,
+                0.883,
+                0.463
+            ],
+            "angle": 0,
+            "content": "Amy Beth Warriner, Victor Kuperman, and Marc Brysbaert. 2013. Norms of valence, arousal, and dominance for 13,915 english lemmas. Behavior research methods, 45(4):1191-1207."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.472,
+                0.883,
+                0.526
+            ],
+            "angle": 0,
+            "content": "Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural text generation with unlikelihood training. arXiv preprint arXiv:1908.04319."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.535,
+                0.883,
+                0.588
+            ],
+            "angle": 0,
+            "content": "John Wieting, Taylor Berg-Kirkpatrick, Kevin Gimpel, and Graham Neubig. 2019. Beyond bleu: training neural machine translation with semantic similarity. arXiv preprint arXiv:1909.06694."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.597,
+                0.883,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.698,
+                0.883,
+                0.738
+            ],
+            "angle": 0,
+            "content": "Yi Zhang, Tao Ge, and Xu Sun. 2020. Parallel data augmentation for formality style transfer. arXiv preprint arXiv:2005.07522."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.738
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.75,
+                0.724,
+                0.765
+            ],
+            "angle": 0,
+            "content": "A Dataset Information"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.775,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We choose four micro-styles from both intended and unintended style categories, based on the style hierarchy as defined in Troiano et al. - Formality, Arousal, Sentiment, and Bias. While formality is considered a \"non-targeted intended\" micro-style, arousal and sentiment are \"targeted intended\" micro-styles. We also include subjective bias, an \"unintended\" micro-style, to ensure we include styles from all hierarchy branches. We built"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3938"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.149
+            ],
+            "angle": 0,
+            "content": "our micro-style joint classification and style transfer models from multiple publicly available NLP datasets built by other researchers, and we detail these below."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.169,
+                0.49,
+                0.298
+            ],
+            "angle": 0,
+            "content": "Formality. We use Grammarly's Yahoo Answers Formality Corpus (Rao and Tetreault, 2018), which consists of \\(105\\mathrm{k}\\) sentences from two styles: \"formal\" and \"informal\" sentences written either in formal or informal modern English. Unlike formal sentences, informal sentences tend to have more misspellings, short forms (\"u\" instead of \"you\"), and non-standard usage of punctuation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.318,
+                0.489,
+                0.414
+            ],
+            "angle": 0,
+            "content": "Arousal. We use the emotion annotated Emobank dataset (Buechel and Hahn, 2022) based on the three-dimensional VAD model developed by (Warriner et al., 2013). In particular, we transform the Arousal dimension into binary categories such as \"arousal\" and \"non-arousal.\""
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.433,
+                0.49,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Sentiment. We use the famous Sentiment140 dataset (Go et al., 2009), which consists of automatically annotated tweets, where the tweets containing positive emoticons are assumed as positive. In contrast, those with negative emoticons are assumed to be negative. The training dataset consisted of 1.6M tweets, and the test dataset consisted of 359 tweets. The tweets were preprocessed using NLTK to remove special Twitter-specific symbols like hashtags, usernames, and URLs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.614,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Bias. We use the Wiki Neutrality Corpus by (Pryzant et al., 2020). This is a new parallel corpus of 180,000 biased and neutralized sentence pairs. In order to train our joint classifier models, we used the training dataset from the appropriate micro-style datasets mentioned above. To implement our style distribution hypothesis, we used random samples for training and testing, from the combination of all the dev datasets from the benchmarks corpus by (Kang and Hovy, 2019). This consists of 15 different styles coupled to both content and domain by varying degrees. We wanted to ensure that the dataset used for training our style transfer model and verifying our hypothesis has sufficient indicators of the appropriate micro-styles. This could be done best by using a sample consisting of datasets curated for each individual micro-style (since a jointly annotated dataset with so many styles is not available)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.799,
+                0.102
+            ],
+            "angle": 0,
+            "content": "B Multi Style Transfer Pipeline"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.123,
+                0.782,
+                0.14
+            ],
+            "angle": 0,
+            "content": "B.1 Resources used for Training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.155,
+                0.885,
+                0.252
+            ],
+            "angle": 0,
+            "content": "All models were trained using cloud GPUs on Google Colab Pro and Pro+. We used 1 V100 GPU in its \"High-RAM\" (52GB) GPU run-time setting to train the paraphrase generation model, while for other models we used 1 P100 GPU at the \"standard RAM\" setting (32GB)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.281,
+                0.807,
+                0.296
+            ],
+            "angle": 0,
+            "content": "B.2 Diverse Paraphrase Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.312,
+                0.885,
+                0.668
+            ],
+            "angle": 0,
+            "content": "We adopted a pre-trained T5 model (Raffel et al., 2020), to generate paraphrases. We trained the model on the ParaNMT-filtered dataset provided by (Krishna et al., 2020). This is a subset of the ParaNMT dataset with filters applied to promote lexical diversity, syntactic diversity, and semantic similarity. This model was then used to generate the pseudo-parallel training data for transfer. We selected the t5-small architecture (60 million parameters) as this is approximately 10x smaller than the GPT-2 large model used in (Krishna et al., 2020). We used the hyper-parameters given in Table 4. Based on the recommendation in the appendix of Raffel et al, we used the \"paraphrase:\" prefix to train the paraphrase model. Once we had this trained paraphrase model, we used diverse beam search (Vijayakumar et al., 2016) to generate diverse paraphrased outputs. The hyper-parameters used for diverse beam search are mentioned in Table 5. We preferred beam search over top-p sampling in order to prioritize fluent paraphrases (Welleck et al., 2019) over unique paraphrases."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.688,
+                0.883,
+                0.72
+            ],
+            "angle": 0,
+            "content": "Input - paraphrase: I love to play my guitar and I do not know why"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.72,
+                0.883,
+                0.752
+            ],
+            "angle": 0,
+            "content": "Output - I love playing my guitar and I'm not sure why"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.532,
+                0.775,
+                0.864,
+                0.877
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>batch size</td><td>8</td></tr><tr><td>number of epochs</td><td>12</td></tr><tr><td>learning rate</td><td>1e-4</td></tr><tr><td>max sequence length</td><td>64</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.886,
+                0.885,
+                0.915
+            ],
+            "angle": 0,
+            "content": "Table 4: Hyper parameters for T5 training for paraphrase generation"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3939"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.137,
+                0.082,
+                0.468,
+                0.215
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>max length</td><td>70</td></tr><tr><td>early stopping</td><td>True</td></tr><tr><td>no repeat ngram size</td><td>5</td></tr><tr><td>num beams</td><td>9</td></tr><tr><td>num beam groups</td><td>3</td></tr><tr><td>diversity penalty</td><td>0.5</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.152,
+                0.224,
+                0.45,
+                0.24
+            ],
+            "angle": 0,
+            "content": "Table 5: Hyperparameters for Beam Search"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.258,
+                0.364,
+                0.274
+            ],
+            "angle": 0,
+            "content": "B.3 Micro-style Classification"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.279,
+                0.49,
+                0.714
+            ],
+            "angle": 0,
+            "content": "We trained a joint sentence classification model to classify the sentence on multiple axes inspired by the approach in Kang and Hovy, which uses an encoder-decoder-based model that learns cross-style patterns with the shared internal representation across styles. Our joint model comprises fully connected layers attached to a DistilBERT model (Sanh et al., 2019), which acts as an encoder. The hyperparameters for this joint model are given in Table 6. This single model effectively replaces the need for a different model for each classification task, significantly reducing the need for computing resources for training and inference. Our joint classifier is essential for downstream tasks like training style transfer models and evaluation. Say we first perform a joint classification of both formality and arousal micro-styles on our datasets, considering we want a multiple-style transfer along the axes of formality and arousal. This results in both formality and arousal pseudo-labels for the sentences. Since these labels are generated algorithmically rather than by hand, we refer to them as pseudo-labels. Pseudo-labeled sentences can then be used to generate the pseudo-parallel dataset for training joint style transfer models and directly measure the variation of a style along the axis of interest."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.137,
+                0.739,
+                0.468,
+                0.824
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>train batch size</td><td>256</td></tr><tr><td>test batch size</td><td>512</td></tr><tr><td>number of epochs</td><td>3</td></tr><tr><td>learning rate</td><td>1e-4</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.146,
+                0.833,
+                0.456,
+                0.849
+            ],
+            "angle": 0,
+            "content": "Table 6: Hyperparameters for Joint Classifier"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.867,
+                0.422,
+                0.882
+            ],
+            "angle": 0,
+            "content": "B.4 Pseudo Parallel Data Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.888,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We then selected the best paraphrase (most stylistically different from the anchor sentence) based"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.196
+            ],
+            "angle": 0,
+            "content": "on the cosine distance between the anchor and the paraphrased sentence's style vectors. To enable the transfer model to transfer to specified levels of a particular style, we defined 'very low', 'low', 'mid', 'high', and 'very high' buckets for each micro-style. In the following, we describe the bucket boundaries for our style scores."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.198,
+                0.884,
+                0.246
+            ],
+            "angle": 0,
+            "content": "Buckets: Very Low = [0, 0.2] Low = [0.2, 0.4]  \nMid = [0.4, 0.6] High = [0.6, 0.95] Very High = [0.95,1]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.247,
+                0.885,
+                0.601
+            ],
+            "angle": 0,
+            "content": "Using the absolute difference between original text style scores and their best-paraphrased text style scores, we find paraphrasing successfully stripped away both formality and arousal aspects of the text. The same phenomenon has been observed in previous studies, such as (Krishna et al., 2020). To ensure a diverse pseudo-parallel dataset, we retain only anchor-paraphrase pairs that do not match in terms of their style bucket. For example, if an anchor-paraphrase sentence pair is assigned style buckets for formality and arousal, as [very high, low] and [very high, very low], this pair will be retained. However, if both style buckets match, the sentence pair will not be considered diverse enough to remain in the pseudo-parallel dataset. In style transfer models, the paraphrased sentence and its style buckets are used as input parameters, while the style buckets assigned to the anchor sentence are used as proxy levels of output style transfer. The following is an example of the input to the joint style transfer model and the expected output."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.617,
+                0.885,
+                0.696
+            ],
+            "angle": 0,
+            "content": "Goal - Highly increase the formality of the sentence, slightly increase the arousal of the sentence  \nInput - transfer: I'm sad you're going | input formality: low | input arousal: low | output formality: high | output arousal: mid"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.698,
+                0.813,
+                0.714
+            ],
+            "angle": 0,
+            "content": "Output - I am sorry you are going to go."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.532,
+                0.726,
+                0.864,
+                0.812
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>train batch size</td><td>8</td></tr><tr><td>test batch size</td><td>8</td></tr><tr><td>number of epochs</td><td>5</td></tr><tr><td>learning rate</td><td>1e-4</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.521,
+                0.821,
+                0.871,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Table 7: Hyperparameters for T5 for Style Transfer"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.867,
+                0.744,
+                0.882
+            ],
+            "angle": 0,
+            "content": "B.5 Style Transfer Training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our T5 models were trained on pseudo-parallel datasets created and filtered as described earlier."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3940"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.192,
+                0.083,
+                0.805,
+                0.277
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.289,
+                0.885,
+                0.349
+            ],
+            "angle": 0,
+            "content": "Figure 5: Considering the micro-style combinations such that, Formality [formal = f, informal = i], Bias [biased = b, unbiased = u], Arousal [aroused = e, un-aroused = u], and Sentiment [negative = n, positive = p]; we observe that the micro-style combinations that are rarer have more representation in the \"balanced\" setting than the \"skewed\" setting. The categories fbnp, funp and iben have more representation for balanced setting vs skewed setting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.372,
+                0.49,
+                0.566
+            ],
+            "angle": 0,
+            "content": "According to the task, we converted the classifier predictions into buckets of style based on the chosen style of the original and then paraphrased sentences. To achieve the necessary intensity transfers, we appended this information to the paraphrased sentence, as motivated by the original T5 paper (Raffel et al., 2020). Hyperparameters are mentioned in Table 7. For generating sentences from our trained models, we used a combination of both beam search (Vijayakumar et al., 2016) and nucleus sampling (Holtzman et al., 2019) and chose the top 3 sentences from the generations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.578,
+                0.367,
+                0.594
+            ],
+            "angle": 0,
+            "content": "C Some Additional Results"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.606,
+                0.487,
+                0.636
+            ],
+            "angle": 0,
+            "content": "C.1 Impact of Fluency filter on input training data"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.644,
+                0.49,
+                0.804
+            ],
+            "angle": 0,
+            "content": "We find that filtering the original dataset based on fluency metrics always results in better style transferred output as compared to the transferred output when the input dataset is not filtered. This is intuitive, as better quality input prevents confusion in the style transfer model and leads to better quality output. As a result of this finding, we use a fluency filter (adversarial classifier \\(>0.1\\) and perplexity \\(< 365\\)), before we conduct any of the rest of our experiments with micro-style distributions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.817,
+                0.462,
+                0.849
+            ],
+            "angle": 0,
+            "content": "C.2 Balancing effect on lesser represented style combinations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.856,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In Figure 6, we consider the case where we examine Formality [formal = f, informal = i] and Arousal [aroused = e, un-aroused = u] micro-styles and compare the percentage of specific style combinations"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.523,
+                0.369,
+                0.871,
+                0.48
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.491,
+                0.883,
+                0.577
+            ],
+            "angle": 0,
+            "content": "Figure 6: Considering the micro-style combinations such that, Formality [formal = f, informal = i] and Arousal [aroused = e, un-aroused = u]; we observe that the micro-style combinations that are rare (ie, in) have more representation in the \"balanced\" setting than the \"skewed\" setting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.602,
+                0.884,
+                0.68
+            ],
+            "angle": 0,
+            "content": "in the test sample. We observe that as the number of micro styles increases, a balanced joint distribution leads to more representation in combinations that are less likely to occur such as in or 'informal and neutral'."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.682,
+                0.885,
+                0.858
+            ],
+            "angle": 0,
+            "content": "Figure 5 shows a similarly pronounced effect. Here the number of micro styles is increased, and we can observe that the balanced setting shows higher representation than the skewed setting. An example of an unlikely style combination is fbnp, or \"formal biased neutral and positive\". We also observe that as the number of micro-style increases, there is no representation in some combinations in both settings [ibnp, iunn, iuen, iunp]. This is a natural result as some micro-style combinations cannot exist in nature."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3941"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.133,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.127,
+                0.533,
+                0.158
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work? 5"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.169,
+                0.554,
+                0.201
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work? 5"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.212,
+                0.697,
+                0.244
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.256,
+                0.67,
+                0.289
+            ],
+            "angle": 0,
+            "content": "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.13,
+                0.127,
+                0.697,
+                0.289
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.299,
+                0.49,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.322,
+                0.161,
+                0.335
+            ],
+            "angle": 0,
+            "content": "2,3"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.347,
+                0.531,
+                0.378
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used? 2,3"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.391,
+                0.779,
+                0.423
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.434,
+                0.882,
+                0.514
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.525,
+                0.882,
+                0.589
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.6,
+                0.882,
+                0.648
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.658,
+                0.882,
+                0.753
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.347,
+                0.882,
+                0.753
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.765,
+                0.494,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.787,
+                0.186,
+                0.801
+            ],
+            "angle": 0,
+            "content": "3,B, C"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.812,
+                0.882,
+                0.86
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? B.1"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.867,
+                0.878,
+                0.89
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3942"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.084,
+                0.88,
+                0.116
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.118,
+                0.167,
+                0.13
+            ],
+            "angle": 0,
+            "content": "B"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.143,
+                0.882,
+                0.191
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.193,
+                0.167,
+                0.205
+            ],
+            "angle": 0,
+            "content": "B"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.218,
+                0.882,
+                0.266
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.269,
+                0.167,
+                0.281
+            ],
+            "angle": 0,
+            "content": "A"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.294,
+                0.878,
+                0.31
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.315,
+                0.215,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.341,
+                0.882,
+                0.373
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.374,
+                0.351,
+                0.389
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.4,
+                0.882,
+                0.448
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.449,
+                0.351,
+                0.464
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.476,
+                0.882,
+                0.523
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.524,
+                0.351,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.55,
+                0.875,
+                0.565
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.567,
+                0.351,
+                0.582
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.593,
+                0.882,
+                0.624
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.626,
+                0.351,
+                0.642
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3943"
+        }
+    ]
+]

2023/Balancing the Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer/e52f8ce0-4325-4d7a-9b9b-68ca012cb308_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12fa759d3aca63f35808467e85c1245f4d0bc651c1eab99279c19acc4a93b85c
+size 761284

2023/Balancing the Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer/full.md

@@ -0,0 +1,323 @@
+# Balancing Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer
+
+Debarati Das David Ma Dongyeop Kang
+
+Department of Computer Science, University of Minnesota
+
+{das00015, maxxx818, dongyeop}@umn.edu
+
+# Abstract
+
+Text style transfer is an exciting task within the field of natural language generation that is often plagued by the need for high-quality paired datasets. Furthermore, training a model for multi-attribute text style transfer requires datasets with sufficient support across all combinations of the considered stylistic attributes, adding to the challenges of training a style transfer model. This paper explores the impact of training data input diversity on the quality of the generated text from the multi-style transfer model. We construct a pseudo-parallel dataset by devising heuristics to adjust the style distribution in the training samples. We balance our training dataset using marginal and joint distributions to train our style transfer models. We observe that a balanced dataset produces more effective control effects over multiple styles than an imbalanced or skewed one. Through quantitative analysis, we explore the impact of multiple style distributions in training data on style-transferred output. These findings will better inform the design of style-transfer datasets.
+
+# 1 Introduction
+
+Multi-style text transfer is a challenging task today with applications such as automatic domain-appropriate, style-conformant writing (Fu et al., 2018) and AI-assisted stylistic language editing. Text style transfer is an intricate task as all language has a specific context, and those contexts influence the attributes of the language (Hovy and Yang, 2021). Text style transfer is challenging because it involves dealing with the aspects of style coupled with the textual content (Hu et al., 2017; Shen et al., 2017; Lample et al., 2018). This domain's other obstacles include the need for parallel corpus (Jhamtani et al., 2017) and quality training data. As the number of style dimensions increases with multi-style text transfer, not only is the requirement of a jointly annotated corpus across all the
+
+![](images/2794bbad8e426aa268b19892248a0e6d7b2378e8b55fe42f584182a0884f29f0.jpg)  
+Figure 1: When an input sentence is passed to the multi-style transfer model, to increase formality and decrease arousal, we hypothesize that when the model is trained on a balanced joint distribution of formality and arousal (all four style combinations have a  $25\%$  representation) - the style transfer is more successful as opposed to when the model is trained on a skewed joint distribution (there is no representation of the "informal unaroused" style combination) of styles in the training data.
+
+stylistic dimensions problematic, but the different styles are not necessarily independent.
+
+While "style" can also refer to authorial or domain-specific style, in this paper, we focus on "microstyles" as defined by (Kang and Hovy, 2021) where they define "micro-style" as a complex combination of different factors such as formality markers, emotions, and metaphors. People intentionally (Troiano et al., 2021) tune these styles in writing differently based on their mood, the person they are addressing, the content of the message, or the platform. Multiple microstyles can jointly describe a text; for example, a given text could simultaneously be formal and sad. Microstyles also more easily lend themselves to being represented as spectra with varying degrees of intensity. These points align with our vision of an application where users can edit micro-style aspects of their writing.
+
+Much research exists on models implementing multi-style text transfer and interdependency of micro-styles (Kang and Hovy, 2019; Goyal et al.,
+
+![](images/8fb69b8365f1aa2b8a81111b213cfabf68d9acec2e6d93d651dc5ce7fa6ff820.jpg)  
+Figure 2: The input sentence transitions through every step in our multi-style text style transfer pipeline. The box in red indicates our main contribution to the pipeline, which helps us explore the effects of joint micro-style combinations on style-transferred output.
+
+2020; Subramanian et al., 2018). However, there needs to be more exploration of the joint distribution of inherent micro-styles in the style transfer training dataset and how these micro-style distributions are related. Therefore, we pose a question - Can a dataset with minimal variance across multiple micro-style combinations, such that it experiences a "balancing effect", lead to a better style transferred output? Figure 1 illustrates our intuition that a dataset that experiences a "balancing effect" will have more control over the multi-style transferred output than a "skewed" dataset. Suppose the style transfer model sees examples of every style combination that can exist - this could aid in the style generation of even unlikely combinations of styles compared to a skewed distribution of these joint micro-styles.
+
+In this research, we consider a multi-style text style transfer pipeline assuming that the user has no access to parallel data or the style of the original text that he wishes to transfer, as would seem natural for a style language editing application. We introduce the changing of the training dataset micro-style joint distributions in such a pipeline and quantitatively explore the impact of this modification on the style transferred output. We perform a set of empirical analyses to demonstrate the influence of joint distributions on style-transferred output and show how this trend varies as the number of micro-styles considered changes. The 'balancing effect' on a training dataset leads to style transferred sentences from even the joint style combinations that are typically rare ("informal unbiased and
+
+unaroused"). Our study is the first of its kind on the distribution of micro styles in training datasets for multi-style text style transfer and is likely to have implications for designing datasets for multi-style transfer model training and fall within the context of and align with recent work on characterizing datasets and factors impacting style transfer (Bender and Friedman, 2018; Schoch et al., 2021; Li et al., 2019; Zhang et al., 2020; Gururangan et al., 2018).
+
+# 2 Multi Style Transfer Pipeline
+
+Datasets: We chose four micro-styles from the style hierarchy defined in Troiano et al.: Formality, Arousal, Sentiment, and Bias, for our study and used publicly available NLP datasets built by other researchers (Rao and Tetreault, 2018; Buechel and Hahn, 2022; Go et al., 2009; Pryzant et al., 2020; Kang and Hovy, 2019) to develop and test our models. Appendix A mentions the details of the datasets and their usage.
+
+Pipeline Overview: Our experimental setup for multi-style transfer is inspired by the work of (Krishna et al., 2020). Like them, we first generate a "diverse" paraphrase of the input sentence, and then the paraphrased sentence is rewritten in the style of choice. Towards this end, we train a paraphrasing model (separately on a parallel paraphrase dataset). Then, the trained paraphrase model is used to create "pseudo-gold" parallel data for training style models.
+
+First, we adopted a pre-trained T5 model (Raffel et al., 2020) to generate paraphrases. This model
+
+was trained for the task of paraphrase generation on the ParaNMT-filtered dataset provided by (Krishna et al., 2020). Once we had this trained paraphrase model, we used diverse beam search (Vijayakumar et al., 2016) to generate diverse fluent paraphrased outputs. An important assumption is that the paraphrase is stripped of its original style and does not leak into the training.
+
+We address this potential issue by training classifiers (Sanh et al., 2019) to predict style on the original and paraphrased datasets and find that all our micro-style classifiers have a classification accuracy of higher than  $80\%$  F1, which is acceptable for pseudo-label creation. After we generate diverse paraphrases, we choose the most diverse paraphrase and then derive micro-style classifications for the paraphrased sentence using our trained micro-style classifiers. Therefore each sentence is assigned a classification score for each micro-style label and can form a "pseudo parallel" dataset for training the T5-based joint transfer model. Thus, our approach does not need a parallel dataset.
+
+We then converted the classifier predictions into buckets of style (ranging from "very low" to "very high") based on the chosen style of the original and then paraphrased sentences. The bucketing process is described in Appendix B. After this step, we introduce our contribution of "constructing style distributions" into the pipeline, as illustrated in Figure 2. Following that, we perform multi-style text style transfer. We appended the "bucket" information to the paraphrased sentence to achieve the necessary intensity transfers, as motivated by the original T5 paper (Raffel et al., 2020). We train T5-based style transfer models, where the paraphrased sentence and its style buckets are used as input parameters, while the style buckets assigned to the anchor sentence are used as proxy levels of output style transfer. All model-specific details are provided in Appendix B. For generating sentences from our trained models, we used beam search (Vijayakumar et al., 2016) and nucleus sampling (Holtzman et al., 2019) and chose the top 3 sentences from the generations. The following is an example of the input to the joint style transfer model and the expected output.
+
+Goal - Highly increase the formality of the sentence, slightly increase the arousal of the sentence
+
+Input - transfer: I'm sad you're going | input
+
+formality: low | input arousal: low | output formality: high | output arousal: mid Output - I am sorry you are going to go.
+
+Thus, we implemented a multi-style transfer pipeline to test our hypothesis without any finicky modeling paradigms popular in style transfer research, such as variational inference or autoregressive sampling (He et al., 2020; Subramanian et al., 2018).
+
+<table><tr><td>Style Combination</td><td>Balanced</td><td>Skewed</td></tr><tr><td>Formal Aroused</td><td>3395</td><td>8685</td></tr><tr><td>Formal Unaroused</td><td>3395</td><td>2792</td></tr><tr><td>Informal Aroused</td><td>3395</td><td>1275</td></tr><tr><td>Informal Unaroused</td><td>3395</td><td>828</td></tr></table>
+
+Table 1: Training data statistics (number of samples) for the balanced and skewed settings, when considering the micro-styles of Formality and Arousal.
+
+Constructing Micro-style Distributions We define a "style combination" as a possible combination of the states that the micro-styles can take together - such as 'informal biased negative.' Since there are three micro-styles, each having binary states, the total possible number of style combinations, in this case, is given by  $N_{c} = 2 \times 2 \times 2 = 2^{3}$ . Therefore to generalize, if  $|m_{i}|$  indicates the cardinality of each micro-style and  $n$  indicates the number of micro-styles considered, the total possible number of style combinations  $(N_{c})$  possible is given by:
+
+$$
+N _ {c} = \prod_ {i = 1} ^ {n} | m _ {i} | \tag {1}
+$$
+
+To create the balanced joint distribution of styles, we ensure the standard deviation across the style combinations is close to 0. We do this by down-sampling each style combination, such that the number of samples in each style combination is the same as the least represented style combination. As we increase micro-styles, some micro-style combinations do not occur naturally together, so their representation is close to 0. In such cases, we assume that the least represented style combination is at least  $5\%$  of the total dataset. To ensure our comparison across the "balanced" and "skew" settings is fair, we construct a skewed dataset with a total sample size that is the same as that of the balanced dataset. Thus, the balanced dataset has a uniform distribution, while the skewed dataset
+
+has a non-uniform distribution. Table 1 shows the number of samples in each style combination of Formality and Arousal, given a "balanced" and "skewed" setting.
+
+# 3 Experimental Results and Discussion
+
+Evaluation Metrics: Style transfer accuracy metrics quantify how nicely output texts match the desired style. However, more than this metric is required. Motivated by Jin et al., we evaluate style transfer across the three main properties of text style transfer: style transfer accuracy, content preservation, and fluency. We use our custom joint sequence classification models, implemented using HuggingFace libraries (Wolf et al., 2020) to evaluate the style transfer success ratio. Our definition for the Style Transfer Success  $S_{c}$  is the total number of matches between intended and transferred style buckets, divided by the total number of samples. To judge content preserved in style transferred text, we use three metrics: BLEU (Papineni et al., 2002), embedding-based similarity (Wieting et al., 2019) using cosine similarity of two sentence embeddings (Reimers and Gurevych, 2019), and Word Mover's Distance (WMD) (Mir et al., 2019). For fluency, we use measures like perplexity using GPT2 (Radford et al., 2019) and an adversarial classifier using the cross-aligned autoencoder model (Mir et al., 2019).
+
+Experimental Setup: In this paper, we illustrate different micro-style combinations in the training data, for a randomly selected case, with each combination in both the "balanced" and "skewed settings. Therefore, we consider 6 cases respectively: 1) Formality and Arousal in a balanced setting (FA balanced) 2) Formality and Arousal in a skewed setting (FA skewed) 3) Formality, Arousal and Bias in a balanced setting (FAB balanced) 4) Formality, Arousal and Bias in skewed setting (FAB skewed) 5) Formality, Arousal, Bias and Sentiment in the balanced setting (FABS balanced) 6) Formality, Arousal, Bias and Sentiment in skewed setting (FABS skewed). We construct the training data with the appropriate settings and then pass them through our experimental pipeline (illustrated in Figure 2) and quantitatively evaluate the style transfer results.
+
+Discussion: Table 2 shows examples of style-transferred sentences, given a style-transfer goal from our experimental pipeline for both balanced and skewed settings. E.g., given the objective is to decrease Formality but increase arousal, the sen
+
+![](images/1e9d114957a6fb214469fb38dffc5def353c5f7ec421f695d58efbfe4690f83e.jpg)  
+Figure 3: Balancing micro-style distributions leads to a higher multi-style transfer percentage than in the Skewed setting in all the cases.
+
+![](images/34aaafb919848ec80a940e533a924f4347a57f86c667d381a9bd298128a21add.jpg)  
+Figure 4: Considering the micro-style combinations such that, Formality [formal = f, informal = i], Bias [biased = b, unbiased = u], and Arousal [aroused = e, un-aroused = n], we observe that the micro-style combinations that are rarer (e.g., informal unbiased neutral (iun)) have more representation in the "balanced" setting than the "skewed" setting.
+
+tence “Did you hear about the soldier with 8 limbs? He was army” transforms to “He’s an army soldier with 8 legs?”. Here, the contraction “He’s” indicates a formality decrease, and the replacement of limbs with legs indicates a decrease. The overall arousal of this sentence is higher when it transforms into a question.
+
+Figure 3 illustrates that the balanced setup always has a higher success percentage of style transfer  $(S_{c})$  than the skewed setup. We cannot compare the success percentage across cases because matching the exact target and transferred style buckets becomes difficult as the number of micro-styles increases. We can also observe through Table 2 that the quality of the balanced transferred text aligns better with the style transfer goal than the skewed transferred text.
+
+In Figure 4, we compare the difference in representation percentage of specific style combinations in the test sample for a specific case where we consider Formality, Arousal, and Bias micro-styles. We observe that a balanced joint distribution leads to more representation in the style combinations that are less likely to occur. This is further accentuated as micro-styles increase, as reported in
+
+<table><tr><td>Style Transfer Goal</td><td>Input Text</td><td>Balanced Transferred Text</td><td>Skewed Transferred Text</td></tr><tr><td>↑ Formality ↑ Arousal</td><td>Wouldn’t it be great if Trump went 3rd party and sucked away millions of Republican votes lol</td><td>It wouldn’t be nice if Trump went to the third party and swooped millions of Repub-lican votes.</td><td>Would it not be great if Trump went 3rd party and sucked away millions of Republican votes?</td></tr><tr><td>↑ Formality ↓ Arousal</td><td>I didn’t know what happiness was until I got married. But by then it was too late.</td><td>Until I got married, I didn’t even know what happiness was.</td><td>I did not know what happiness was till I got married and it was too late.</td></tr><tr><td>↓ Formality ↑ Arousal</td><td>Did you hear about the soldier with 8 limbs? He was army</td><td>He’s an army soldier with 8 legs?</td><td>Did you hear about the soldier with 8 limbs in the army?</td></tr><tr><td>↓ Formality ↓ Arousal</td><td>Yeah, I don’t understand all the hate.</td><td>Yeah I’m not gonna understand the hate.</td><td>Yeah I do not understand all the hate.</td></tr></table>
+
+Appendix C. In Figure 4, we see that rarer style combinations [ibn, fun, iun] show more representation in the balanced case as compared to the skewed case. This supports our intuition that the style transfer model benefits from learning the representation of all possible style combinations that can occur together.
+
+When we consider Formality, Arousal, and Bias micro styles together, the most represented category (30% of samples) is "formal unbiased aroused" (fue). The least represented category (as unlikely to occur together) is "informal unbiased unaroused" (iun) with 1%. We observe that the quantitative evaluation metrics are quite indicative when compared across style combinations. For instance, in Table 3, we observe that perplexity increases in categories that are unlikely to occur together (iun). This indicates that the style transfer model is confused by the style distributions present for this style combination.
+
+We do not claim that our method of balancing multiple styles will work even for entangled micro-style combinations, as that is out of the scope of the current paper. However, balancing considerably affects the multi-style transfer output for the range of micro-style combinations we considered, and that has an application in many NLP tasks. This result could hugely influence future studies exploring better ways to balance even the entangled micro-styles.
+
+# 4 Conclusion
+
+Multi-style text style transfer is a challenging problem predominantly plagued by the need for jointly annotated high-quality datasets. There is a clear need for more research about the marginal and joint distribution of inherent micro-styles present in the training dataset used for style transfer. Multi-style text-style transfer typically requires access to large, jointly labeled datasets and many computational resources under typical implementations. More
+
+Table 2: The table shows the style transferred sentences, given an input sentence and the intended style transfer goal, for both the balanced setting as well as the skewed setting.  
+
+<table><tr><td>Setting</td><td>Styles</td><td>Perp</td><td>Adv</td><td>BLEU</td><td>Cos</td><td>WMD</td></tr><tr><td rowspan="2">Balanced</td><td>fue</td><td>115.16</td><td>0.90</td><td>0.77</td><td>0.92</td><td>0.32</td></tr><tr><td>iun*</td><td>598.58</td><td>0.86</td><td>0.78</td><td>0.92</td><td>0.36</td></tr><tr><td rowspan="2">Skewed</td><td>fue</td><td>116.02</td><td>0.90</td><td>0.78</td><td>0.92</td><td>0.32</td></tr><tr><td>iun*</td><td>650.47</td><td>0.82</td><td>0.77</td><td>0.93</td><td>0.37</td></tr></table>
+
+Table 3: Comparison of the evaluation metrics for the most represented style combination (fue - formal unbiased aroused) vs the least represented style combination (iun* - informal unbiased unaroused). One key observation is that perplexity increases when the style combinations are unlikely to occur together.
+
+importantly, we would not be able to conveniently tweak the input data distributions in other multi-style text style transfer methods.
+
+In this paper, we implement a multi-style transfer pipeline that subverts the requirement of a jointly annotated dataset of multiple styles by constructing a pseudo-parallel dataset to which we introduce our contribution of constructing style distributions. We then use the modified pseudo-parallel datasets for multi-style transfer. Our modified pipeline effectively allows us to understand the importance of the joint distribution of micro styles in training data and is a substantial contribution.
+
+We quantitatively explore the impact of joint micro-style distributions in the training dataset on the style-transferred output sentences. When the joint micro-style distributions are balanced, there is more control over style-transferred output than with a skewed distribution. These findings will likely inform the design of multi-style transfer datasets and encourage us to explore the micro-style relationships in our datasets.
+
+# Limitations
+
+In this research, though we employed automatic evaluation of our multi-style transferred text, we acknowledge that multi-style transfer is challenging to observe with the existing metrics for style transfer evaluation, and human evaluation should
+
+be done as well. As this research paper focuses on exploring the impact of style distributions in the training data on style-transferred output rather than developing a superior multi-style text transfer model, we use quantitative evaluation in this iteration of our paper. We hope that the large sample size and the consistency of applied metrics make our automated approach a reasonable way of evaluating the style transfer output.
+
+This iteration of our paper aims to achieve multi-style transfer across multiple micro styles taken into consideration together as our contribution would aid in constructing a training dataset for multiple micro-style style transfers. We did not explore another exciting question of how balancing multiple micro styles in the training dataset might influence individual style transfer, which could be a promising future direction for our study.
+
+We acknowledge that the classifier's quality sets an upper bound on the best style transfer accuracy that is obtainable. However, the target task is quite complicated without a parallel dataset. Our objective was not to have the most accurate classification of micro styles but to find a means to get acceptable pseudo labels for the micro styles. Individually, all our micro style classifiers had a classification accuracy of  $80\%$  F1 and higher, and we deemed this good enough for pseudo-label creation.
+
+We also focused on utilizing the present styles in the training data and classifying them to derive inherent training style distributions instead of dynamically tuning the proportion of styles present in the training dataset. However, tuning these style proportions using techniques such as PPLM (Dathathri et al., 2019) would give us greater control over our experimental pipeline and is an appropriate next step.
+
+# Acknowledgement
+
+We thank Vivek Aithal, Priyam Srivastava and Daniel McAndrew for their initial work on the pipeline for multi-style transfer. This was instrumental to our project and helped us get a kickstart on our research.
+
+# References
+
+Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587-604.
+
+Sven Buechel and Udo Hahn. 2022. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. arXiv preprint arXiv:2205.01996.  
+Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2019. Plug and play language models: A simple approach to controlled text generation. arXiv preprint arXiv:1912.02164.  
+Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Exploration and evaluation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.  
+Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford, 1(12):2009.  
+Navita Goyal, Balaji Vasan Srinivasan, Anandhavelu Natarajan, and Abhilasha Sancheti. 2020. Multi-style transfer with discriminative feedback on disjoint corpus. arXiv preprint arXiv:2010.11578.  
+Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324.  
+Junxian He, Xinyi Wang, Graham Neubig, and Taylor Berg-Kirkpatrick. 2020. A probabilistic formulation of unsupervised text style transfer. arXiv preprint arXiv:2002.03912.  
+Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751.  
+Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In *The 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*. Association for Computational Linguistics.  
+Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In International conference on machine learning, pages 1587-1596. PMLR.  
+Harsh Jhamtani, Varun Gangal, Eduard Hovy, and Eric Nyberg. 2017. Shakespearizing modern language using copy-enriched sequence-to-sequence models. arXiv preprint arXiv:1707.01161.  
+Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea. 2022. Deep learning for text style transfer: A survey. Computational Linguistics, 48(1):155-205.  
+Dongyeop Kang and Eduard Hovy. 2019. xslue: A benchmark and analysis platform for cross-style language understanding and evaluation. arXiv preprint arXiv:1911.03663.
+
+Dongyeop Kang and Eduard Hovy. 2021. Style is not a single variable: Case studies for cross-stylistic language understanding. 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 2376-2387.  
+Kalpesh Krishna, John Wieting, and Mohit Iyyer. 2020. Reformulating unsupervised style transfer as paraphrase generation. arXiv preprint arXiv:2010.05700.  
+Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y-Lan Boureau. 2018. Multiple-attribute text rewriting. In International Conference on Learning Representations.  
+Dianqi Li, Yizhe Zhang, Zhe Gan, Yu Cheng, Chris Brockett, Ming-Ting Sun, and Bill Dolan. 2019. Domain adaptive text style transfer. arXiv preprint arXiv:1908.09395.  
+Remi Mir, Bjarke Felbo, Nick Obradovich, and Iyad Rahwan. 2019. Evaluating style transfer for text. CoRR, abs/1904.02295.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, and Diyi Yang. 2020. Automatically neutralizing subjective bias in text. In Proceedings of the aaai conference on artificial intelligence, volume 34, pages 480-489.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.  
+Sudha Rao and Joel Tetreault. 2018. Dear sir or madam, may i introduce the gyafc dataset: Corpus, benchmarks and metrics for formality style transfer. arXiv preprint arXiv:1803.06535.  
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks.  
+Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.
+
+Stephanie Schoch, Wanyu Du, and Yangfeng Ji. 2021. Contextualizing variation in text style transfer datasets. arXiv preprint arXiv:2108.07871.  
+Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. Advances in neural information processing systems, 30.  
+Sandeep Subramanian, Guillaume Lample, Eric Michael Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y-Lan Boureau. 2018. Multiple-attribute text style transfer. arXiv preprint arXiv:1811.00552.  
+Enrica Troiano, Aswathy Velutharambath, et al. 2021. From theories on styles to their transfer in text: Bridging the gap with a hierarchical survey. arXiv preprint arXiv:2110.15871.  
+Ashwin K Vijayakumar, Michael Cogswell, Ramprasath R Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424.  
+Amy Beth Warriner, Victor Kuperman, and Marc Brysbaert. 2013. Norms of valence, arousal, and dominance for 13,915 english lemmas. Behavior research methods, 45(4):1191-1207.  
+Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural text generation with unlikelihood training. arXiv preprint arXiv:1908.04319.  
+John Wieting, Taylor Berg-Kirkpatrick, Kevin Gimpel, and Graham Neubig. 2019. Beyond bleu: training neural machine translation with semantic similarity. arXiv preprint arXiv:1909.06694.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45.  
+Yi Zhang, Tao Ge, and Xu Sun. 2020. Parallel data augmentation for formality style transfer. arXiv preprint arXiv:2005.07522.
+
+# A Dataset Information
+
+We choose four micro-styles from both intended and unintended style categories, based on the style hierarchy as defined in Troiano et al. - Formality, Arousal, Sentiment, and Bias. While formality is considered a "non-targeted intended" micro-style, arousal and sentiment are "targeted intended" micro-styles. We also include subjective bias, an "unintended" micro-style, to ensure we include styles from all hierarchy branches. We built
+
+our micro-style joint classification and style transfer models from multiple publicly available NLP datasets built by other researchers, and we detail these below.
+
+Formality. We use Grammarly's Yahoo Answers Formality Corpus (Rao and Tetreault, 2018), which consists of  $105\mathrm{k}$  sentences from two styles: "formal" and "informal" sentences written either in formal or informal modern English. Unlike formal sentences, informal sentences tend to have more misspellings, short forms ("u" instead of "you"), and non-standard usage of punctuation.
+
+Arousal. We use the emotion annotated Emobank dataset (Buechel and Hahn, 2022) based on the three-dimensional VAD model developed by (Warriner et al., 2013). In particular, we transform the Arousal dimension into binary categories such as "arousal" and "non-arousal."
+
+Sentiment. We use the famous Sentiment140 dataset (Go et al., 2009), which consists of automatically annotated tweets, where the tweets containing positive emoticons are assumed as positive. In contrast, those with negative emoticons are assumed to be negative. The training dataset consisted of 1.6M tweets, and the test dataset consisted of 359 tweets. The tweets were preprocessed using NLTK to remove special Twitter-specific symbols like hashtags, usernames, and URLs.
+
+Bias. We use the Wiki Neutrality Corpus by (Pryzant et al., 2020). This is a new parallel corpus of 180,000 biased and neutralized sentence pairs. In order to train our joint classifier models, we used the training dataset from the appropriate micro-style datasets mentioned above. To implement our style distribution hypothesis, we used random samples for training and testing, from the combination of all the dev datasets from the benchmarks corpus by (Kang and Hovy, 2019). This consists of 15 different styles coupled to both content and domain by varying degrees. We wanted to ensure that the dataset used for training our style transfer model and verifying our hypothesis has sufficient indicators of the appropriate micro-styles. This could be done best by using a sample consisting of datasets curated for each individual micro-style (since a jointly annotated dataset with so many styles is not available).
+
+# B Multi Style Transfer Pipeline
+
+# B.1 Resources used for Training
+
+All models were trained using cloud GPUs on Google Colab Pro and Pro+. We used 1 V100 GPU in its "High-RAM" (52GB) GPU run-time setting to train the paraphrase generation model, while for other models we used 1 P100 GPU at the "standard RAM" setting (32GB).
+
+# B.2 Diverse Paraphrase Generation
+
+We adopted a pre-trained T5 model (Raffel et al., 2020), to generate paraphrases. We trained the model on the ParaNMT-filtered dataset provided by (Krishna et al., 2020). This is a subset of the ParaNMT dataset with filters applied to promote lexical diversity, syntactic diversity, and semantic similarity. This model was then used to generate the pseudo-parallel training data for transfer. We selected the t5-small architecture (60 million parameters) as this is approximately 10x smaller than the GPT-2 large model used in (Krishna et al., 2020). We used the hyper-parameters given in Table 4. Based on the recommendation in the appendix of Raffel et al, we used the "paraphrase:" prefix to train the paraphrase model. Once we had this trained paraphrase model, we used diverse beam search (Vijayakumar et al., 2016) to generate diverse paraphrased outputs. The hyper-parameters used for diverse beam search are mentioned in Table 5. We preferred beam search over top-p sampling in order to prioritize fluent paraphrases (Welleck et al., 2019) over unique paraphrases.
+
+Input - paraphrase: I love to play my guitar and I do not know why
+
+Output - I love playing my guitar and I'm not sure why
+
+<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>batch size</td><td>8</td></tr><tr><td>number of epochs</td><td>12</td></tr><tr><td>learning rate</td><td>1e-4</td></tr><tr><td>max sequence length</td><td>64</td></tr></table>
+
+Table 4: Hyper parameters for T5 training for paraphrase generation
+
+<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>max length</td><td>70</td></tr><tr><td>early stopping</td><td>True</td></tr><tr><td>no repeat ngram size</td><td>5</td></tr><tr><td>num beams</td><td>9</td></tr><tr><td>num beam groups</td><td>3</td></tr><tr><td>diversity penalty</td><td>0.5</td></tr></table>
+
+# B.3 Micro-style Classification
+
+We trained a joint sentence classification model to classify the sentence on multiple axes inspired by the approach in Kang and Hovy, which uses an encoder-decoder-based model that learns cross-style patterns with the shared internal representation across styles. Our joint model comprises fully connected layers attached to a DistilBERT model (Sanh et al., 2019), which acts as an encoder. The hyperparameters for this joint model are given in Table 6. This single model effectively replaces the need for a different model for each classification task, significantly reducing the need for computing resources for training and inference. Our joint classifier is essential for downstream tasks like training style transfer models and evaluation. Say we first perform a joint classification of both formality and arousal micro-styles on our datasets, considering we want a multiple-style transfer along the axes of formality and arousal. This results in both formality and arousal pseudo-labels for the sentences. Since these labels are generated algorithmically rather than by hand, we refer to them as pseudo-labels. Pseudo-labeled sentences can then be used to generate the pseudo-parallel dataset for training joint style transfer models and directly measure the variation of a style along the axis of interest.
+
+Table 5: Hyperparameters for Beam Search  
+
+<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>train batch size</td><td>256</td></tr><tr><td>test batch size</td><td>512</td></tr><tr><td>number of epochs</td><td>3</td></tr><tr><td>learning rate</td><td>1e-4</td></tr></table>
+
+# B.4 Pseudo Parallel Data Generation
+
+We then selected the best paraphrase (most stylistically different from the anchor sentence) based
+
+on the cosine distance between the anchor and the paraphrased sentence's style vectors. To enable the transfer model to transfer to specified levels of a particular style, we defined 'very low', 'low', 'mid', 'high', and 'very high' buckets for each micro-style. In the following, we describe the bucket boundaries for our style scores.
+
+Buckets: Very Low = [0, 0.2] Low = [0.2, 0.4]  
+Mid = [0.4, 0.6] High = [0.6, 0.95] Very High = [0.95,1]
+
+Using the absolute difference between original text style scores and their best-paraphrased text style scores, we find paraphrasing successfully stripped away both formality and arousal aspects of the text. The same phenomenon has been observed in previous studies, such as (Krishna et al., 2020). To ensure a diverse pseudo-parallel dataset, we retain only anchor-paraphrase pairs that do not match in terms of their style bucket. For example, if an anchor-paraphrase sentence pair is assigned style buckets for formality and arousal, as [very high, low] and [very high, very low], this pair will be retained. However, if both style buckets match, the sentence pair will not be considered diverse enough to remain in the pseudo-parallel dataset. In style transfer models, the paraphrased sentence and its style buckets are used as input parameters, while the style buckets assigned to the anchor sentence are used as proxy levels of output style transfer. The following is an example of the input to the joint style transfer model and the expected output.
+
+Goal - Highly increase the formality of the sentence, slightly increase the arousal of the sentence  
+Input - transfer: I'm sad you're going | input formality: low | input arousal: low | output formality: high | output arousal: mid
+
+Output - I am sorry you are going to go.
+
+Table 6: Hyperparameters for Joint Classifier  
+
+<table><tr><td>Hyperparameters</td><td>Value</td></tr><tr><td>train batch size</td><td>8</td></tr><tr><td>test batch size</td><td>8</td></tr><tr><td>number of epochs</td><td>5</td></tr><tr><td>learning rate</td><td>1e-4</td></tr></table>
+
+Table 7: Hyperparameters for T5 for Style Transfer
+
+# B.5 Style Transfer Training
+
+Our T5 models were trained on pseudo-parallel datasets created and filtered as described earlier.
+
+![](images/77c2fdd4393c5fcea492bfe16dcec1804e4b64e23d60a0c3a02891931353a5f8.jpg)  
+Figure 5: Considering the micro-style combinations such that, Formality [formal = f, informal = i], Bias [biased = b, unbiased = u], Arousal [aroused = e, un-aroused = u], and Sentiment [negative = n, positive = p]; we observe that the micro-style combinations that are rarer have more representation in the "balanced" setting than the "skewed" setting. The categories fbnp, funp and iben have more representation for balanced setting vs skewed setting.
+
+According to the task, we converted the classifier predictions into buckets of style based on the chosen style of the original and then paraphrased sentences. To achieve the necessary intensity transfers, we appended this information to the paraphrased sentence, as motivated by the original T5 paper (Raffel et al., 2020). Hyperparameters are mentioned in Table 7. For generating sentences from our trained models, we used a combination of both beam search (Vijayakumar et al., 2016) and nucleus sampling (Holtzman et al., 2019) and chose the top 3 sentences from the generations.
+
+# C Some Additional Results
+
+# C.1 Impact of Fluency filter on input training data
+
+We find that filtering the original dataset based on fluency metrics always results in better style transferred output as compared to the transferred output when the input dataset is not filtered. This is intuitive, as better quality input prevents confusion in the style transfer model and leads to better quality output. As a result of this finding, we use a fluency filter (adversarial classifier  $>0.1$  and perplexity  $< 365$ ), before we conduct any of the rest of our experiments with micro-style distributions.
+
+# C.2 Balancing effect on lesser represented style combinations
+
+In Figure 6, we consider the case where we examine Formality [formal = f, informal = i] and Arousal [aroused = e, un-aroused = u] micro-styles and compare the percentage of specific style combinations
+
+![](images/924bd3451ecd9865043c114d2e82e26722acaff2ce9606d5513b4898241e5448.jpg)  
+Figure 6: Considering the micro-style combinations such that, Formality [formal = f, informal = i] and Arousal [aroused = e, un-aroused = u]; we observe that the micro-style combinations that are rare (ie, in) have more representation in the "balanced" setting than the "skewed" setting.
+
+in the test sample. We observe that as the number of micro styles increases, a balanced joint distribution leads to more representation in combinations that are less likely to occur such as in or 'informal and neutral'.
+
+Figure 5 shows a similarly pronounced effect. Here the number of micro styles is increased, and we can observe that the balanced setting shows higher representation than the skewed setting. An example of an unlikely style combination is fbnp, or "formal biased neutral and positive". We also observe that as the number of micro-style increases, there is no representation in some combinations in both settings [ibnp, iunn, iuen, iunp]. This is a natural result as some micro-style combinations cannot exist in nature.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? 5  
+A2. Did you discuss any potential risks of your work? 5  
+A3. Do the abstract and introduction summarize the paper's main claims?  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+2,3
+
+B1. Did you cite the creators of artifacts you used? 2,3  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+C Did you run computational experiments?
+
+3,B, C
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? B.1
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+B
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+B
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+A
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+Not applicable. Left blank.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+Not applicable. Left blank.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+Not applicable. Left blank.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+Not applicable. Left blank.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+Not applicable. Left blank.

2023/Balancing the Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f268d75df631f2039e67263b8c1ff032de114ca44c5a32f1753c98e17678863
+size 311155

2023/BanglaBook_ A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews/28378eaf-0a48-458e-a698-291e5038e7bf_content_list.json

@@ -0,0 +1,1438 @@
+[
+    {
+        "type": "text",
+        "text": "BANGLABOOK: A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews",
+        "text_level": 1,
+        "bbox": [
+            139,
+            82,
+            860,
+            120
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Mohsinul Kabir*, Obayed Bin Mahfuz*, Syed Rifat Raiyan*, Hasan Mahmud, Md Kamrul Hasan",
+        "bbox": [
+            240,
+            130,
+            763,
+            164
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Systems and Software Lab (SSL)",
+        "bbox": [
+            364,
+            165,
+            638,
+            181
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science and Engineering",
+        "bbox": [
+            295,
+            181,
+            707,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Islamic University of Technology, Dhaka, Bangladesh",
+        "bbox": [
+            280,
+            198,
+            719,
+            214
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{mohsinulkabir, siam, rifatraiyan, hasan, hasank}@iut-dhaka.edu",
+        "bbox": [
+            184,
+            215,
+            820,
+            231
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The analysis of consumer sentiment, as expressed through reviews, can provide a wealth of insight regarding the quality of a product. While the study of sentiment analysis has been widely explored in many popular languages, relatively less attention has been given to the Bangla language, mostly due to a lack of relevant data and cross-domain adaptability. To address this limitation, we present BANGLABOOK, a large-scale dataset of Bangla book reviews consisting of 158,065 samples classified into three broad categories: positive, negative, and neutral. We provide a detailed statistical analysis of the dataset and employ a range of machine learning models to establish baselines including SVM, LSTM, and Bangla-BERT. Our findings demonstrate a substantial performance advantage of pretrained models over models that rely on manually crafted features, emphasizing the necessity for additional training resources in this domain. Additionally, we conduct an in-depth error analysis by examining sentiment unigrams, which may provide insight into common classification errors in under-resourced languages like Bangla. Our codes and data are publicly available at https://github.com/mohsinulkabir14/BanglaBook.",
+        "bbox": [
+            144,
+            280,
+            460,
+            677
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            689,
+            258,
+            703
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The resources publicly available for scholarly investigation in the realm of Sentiment Analysis (SA) for the Bangla language are scarce and limited in quantity (Khatun and Rabeya, 2022; Sazzed, 2021; Rahman et al., 2019) despite its literary gravitas as the  $6^{\\text{th}}$  most spoken language $^{1}$  in the world with approximately 200 million speakers. In the existing literature on Bangla Text SA, as shown in Table 5, the largest dataset consists",
+        "bbox": [
+            112,
+            714,
+            489,
+            859
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of 20,468 samples (Islam et al., 2022) while the smallest has a mere 1,050 samples (Tabassum and Khan, 2019). Besides these, Islam et al. (2020) created a dataset consisting of 17,852 samples and Islam et al. (2021) utilized a dataset of 15,728 samples. All other datasets apart from these either have  $< 15,000$  samples or are publicly unavailable. Another limitation of the existing research works in Bangla Text SA is the deficiency of datasets having product-specific review samples. Most of the available Bangla SA datasets are focused on user-generated textual content from cyberspace. The insights derived from these may not accurately represent sentiment in the context of product reviews, thus hindering their usefulness for businesses. The tonal and linguistic analysis of reviews from product-specific datasets can aid businesses to gain valuable insights into customer attitudes, preferences, and experiences which can then be leveraged to improve products and services, design targeted marketing campaigns, and make more informed business decisions. In this paper, we introduce a large-scale dataset, BANGLABOOK, consisting of 158,065 samples of book reviews collected from online bookshops written in the Bangla language. This is the largest dataset for Bangla sentiment analysis to the best of our knowledge. We perform an analysis of the dataset's statistical characteristics, employ various ML techniques to establish a performance benchmark for validating the dataset, and also conduct a thorough evaluation of the classification errors.",
+        "bbox": [
+            507,
+            252,
+            885,
+            766
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            778,
+            727,
+            793
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In order to create this dataset, we collect a total of 204,659 book reviews from two online bookshops (Rokomari² and Wafilife³) using a web scraper developed with several Python libraries, including BeautifulSoup, Selenium, Pandas, Openpyxl,",
+        "bbox": [
+            507,
+            804,
+            884,
+            884
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "$^{2}$ https://www.rokomari.com/",
+            "<sup>3</sup>https://www.wafilife.com/"
+        ],
+        "bbox": [
+            529,
+            890,
+            731,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*These authors contributed equally to this work. Author names are in alphabetic order.",
+        "bbox": [
+            112,
+            866,
+            487,
+            892
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1https://en.wikipedia.org/wiki/List_of_",
+        "bbox": [
+            112,
+            892,
+            433,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "languages_by_total_number_of-speakers",
+        "bbox": [
+            112,
+            904,
+            400,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1237",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 1237-1247",
+        "bbox": [
+            228,
+            945,
+            768,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            700,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/62698ed0ca6e74df8be7fb792b84e37f3a455aae8358ce543d4ad9b2bb9e52c9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Source</td><td>Language</td><td>Annotated</td><td>Unannotated</td><td>Total</td></tr><tr><td rowspan=\"4\">Rokomari</td><td>Bangla</td><td>84,672</td><td>24,806</td><td>109,478</td></tr><tr><td>Bangla† + English</td><td>56,485</td><td>15,408</td><td>71,893</td></tr><tr><td>Bangla + Bangla† + English</td><td>13,144</td><td>4,114</td><td>17,258</td></tr><tr><td>Total</td><td>154,301</td><td>44,328</td><td>198,629</td></tr><tr><td rowspan=\"4\">Wafilife</td><td>Bangla</td><td>4,699</td><td>59</td><td>4,758</td></tr><tr><td>Bangla† + English</td><td>370</td><td>3</td><td>373</td></tr><tr><td>Bangla + Bangla† + English</td><td>896</td><td>3</td><td>899</td></tr><tr><td>Total</td><td>5,965</td><td>65</td><td>6,030</td></tr><tr><td rowspan=\"4\"></td><td>Bangla</td><td>89,371</td><td>24,865</td><td>114,237</td></tr><tr><td>Bangla† + English</td><td>56,855</td><td>15,411</td><td>72,266</td></tr><tr><td>Bangla + Bangla† + English</td><td>14,040</td><td>4,117</td><td>18,157</td></tr><tr><td>Total</td><td>160,266</td><td>44,393</td><td>204,659</td></tr><tr><td></td><td>Untranslated Data (Removed)</td><td>2,201</td><td></td><td></td></tr><tr><td></td><td>Final Dataset Size</td><td>158,065</td><td></td><td></td></tr></table>",
+        "bbox": [
+            273,
+            80,
+            724,
+            234
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "and Webdriver, to collect and process the raw data.",
+        "bbox": [
+            112,
+            266,
+            485,
+            296
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For the data collection and preparation process of the BANGLABOOK dataset, we first compile a list of URLs for authors from online bookstores. From there, we procure URLs for the books. We meticulously scrape information such as book titles, author names, book categories, review texts, reviewer names, review dates, and ratings by utilizing these book URLs.",
+        "bbox": [
+            112,
+            299,
+            489,
+            426
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/9d8c467f457468d5165e196ef021af7e1c68badfd132dbc41fd6caf2a205b023.jpg",
+        "table_caption": [
+            "Table 1: Summary statistics of our dataset. Bangla† denotes Romanized Bangla text."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"5\">General Properties</td><td># of Reviews</td><td>158,065</td></tr><tr><td># of Books</td><td>30,253</td></tr><tr><td># of Reviewers</td><td>44,616</td></tr><tr><td># of Categories</td><td>1,573</td></tr><tr><td>Total Review Words</td><td>44,429,201</td></tr><tr><td rowspan=\"3\">Single Review</td><td>max. Word length</td><td>765</td></tr><tr><td>min. Word length</td><td>1</td></tr><tr><td>avg. Word length</td><td>281.08</td></tr></table>",
+        "bbox": [
+            161,
+            439,
+            442,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 2: General overview of BANGLABOOK.",
+        "bbox": [
+            142,
+            542,
+            455,
+            555
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Labeling & Translation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            575,
+            344,
+            590
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "If a review does not have a rating, we deem it unannotated. Reviews with a rating of 1 or 2 are classified as negative, a rating of 3 is considered neutral, and a rating of 4 or 5 is classified as positive. Two manual experiments are carried out to validate the use of ratings as a measure of sentiment in product reviews. In the first experiment, around  $10\\%$  of the reviews are randomly selected and annotated manually. The annotated labels are cross-checked with the original labels, resulting in a  $96.7\\%$  accuracy in the corresponding labels. In addition, we consult the work of Wang et al. (2020) that explored the issue of incongruous sentiment expressions with regard to ratings. Specifically, the study scrutinized two categories of reviews: high ratings lacking a positive sentiment, and low ratings lacking a negative sentiment. We perform an analysis to identify such inconsistencies within our dataset and discovered that only a minuscule  $3.41\\%$  of the samples exhibited this pattern. This figure is rela",
+        "bbox": [
+            112,
+            596,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tively insignificant when considering the substantially large scale of our dataset.",
+        "bbox": [
+            507,
+            266,
+            882,
+            298
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/76550d660072eace8cce26377b4dcae393302544a68c254e887f398292922d2d.jpg",
+        "image_caption": [
+            "(a) Sentiment Distribution"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            310,
+            678,
+            431
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/21e6fd1cf1fee406816d2a9d897267b82959eb266caa2730ad1e1353cc05df59.jpg",
+        "image_caption": [
+            "(b) Rating Distribution",
+            "Figure 1: Class Distribution of BANGLABOOK."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            680,
+            318,
+            875,
+            432
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "After discarding the unannotated reviews, we curate a final dataset of 158,065 annotated reviews. Of these, 89,371 are written entirely in Bangla. The remaining 68,694 reviews were written in Romanized Bangla, English, or a mix of languages. They are translated into Bangla with Google Translator and a custom Python program using the googletrans library. The translations are subsequently subjected to manual review and scrutiny to confirm their accuracy. The majority of inaccurate translations primarily comprise spelling errors and instances where English words remain untranslated within samples containing a combination of Bangla and English text. The meticulous evaluation process of untranslated samples involves a thorough assessment by post-graduate native Bangla speakers, who critically compare the translated text against the original untranslated text to ascertain the correctness of the translation.",
+        "bbox": [
+            507,
+            475,
+            884,
+            781
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Statistical Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            795,
+            707,
+            812
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Tables 1 and 2 provide an overview of the statistical properties of the BANGLABOOK dataset. The sentiment doughnut chart in Figure-1a illustrates the proportion of positive, neutral, and negative reviews, while the rating doughnut chart in Figure-1b displays the percentage of reviews that corre",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1238",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/6dc46aea45371fb8cc968d77fc94c0d854d073eeb330e7a5a9281f385b6618d6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Negative</td><td>Neutral</td><td>Positive</td><td>Weighted Avg.</td></tr><tr><td>Random Forest (word 2-gram + word 3-gram)</td><td>0.56</td><td>0.34</td><td>0.96</td><td>0.9106</td></tr><tr><td>SVM (word 2-gram + word 3-gram)</td><td>0.40</td><td>0.15</td><td>1.00</td><td>0.9053</td></tr><tr><td>Random Forest (word 1-gram)</td><td>0.48</td><td>0.35</td><td>0.96</td><td>0.9043</td></tr><tr><td>Logistic Regression (char 2-gram + char 3-gram)</td><td>0.55</td><td>0.13</td><td>0.96</td><td>0.8978</td></tr><tr><td>Bangla-BERT (base-uncased)</td><td>0.60</td><td>0.22</td><td>0.96</td><td>0.9064</td></tr><tr><td>Logistic Regression (word 2-gram + word 3-gram)</td><td>0.53</td><td>0.13</td><td>0.96</td><td>0.8964</td></tr><tr><td>Bangla-BERT (large)</td><td>0.72</td><td>0.40</td><td>0.97</td><td>0.9331</td></tr><tr><td>XGBoost (char 2-gram + char 3-gram)</td><td>0.31</td><td>0.02</td><td>0.95</td><td>0.8723</td></tr><tr><td>Multinomial NB (word 2-gram + word 3-gram)</td><td>0.23</td><td>0.03</td><td>0.95</td><td>0.8663</td></tr><tr><td>LSTM (GloVe)</td><td>0.11</td><td>0.00</td><td>0.10</td><td>0.0991</td></tr><tr><td>XGBoost (word 2-gram + word 3-gram)</td><td>0.23</td><td>0.01</td><td>0.95</td><td>0.8651</td></tr><tr><td>Multinomial NB (BoW)</td><td>0.18</td><td>0.05</td><td>0.94</td><td>0.8564</td></tr><tr><td>SVM (word 1-gram)</td><td>0.08</td><td>0.04</td><td>0.94</td><td>0.8519</td></tr></table>",
+        "bbox": [
+            129,
+            80,
+            870,
+            231
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Catergory-wise Binary Task F1-score and Weighted Average F1-score of each method on BANGLABOOK.",
+        "bbox": [
+            112,
+            235,
+            880,
+            250
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "spond to each rating on a scale of 1 to 5.",
+        "bbox": [
+            112,
+            263,
+            413,
+            278
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Upon analyzing the sentiment chart, it appears that the majority of the reviews  $(124,084 + 17,503 = 141,587$  samples) are positive, with a significant portion also being negative  $(2,728 + 6,946 = 9,674$  samples). A relatively small fraction of the reviews are neutral (6,804 samples). This suggests that overall, the books have been well received by the readers, with the majority expressing favorable opinions. The distribution of the dataset is representative of real-world scenarios and it tessellates well with previous content analysis works on book reviews (Lin et al., 2005; Sorensen and Rasmussen, 2004). In Figure-2, we can visualize an illustration of the sentiment distribution among the 5 most frequently reviewed categories of books. We can gain some salient insights from the popularity of these genres. Contemporary novels are bestsellers as they reflect current events, social issues, and trends, making them relatable and thought-provoking for the readers while self-help and religious books provide guidance, inspiration, and a sense of purpose, catering to individuals' quest for personal growth and spiritual fulfillment.",
+        "bbox": [
+            112,
+            279,
+            489,
+            650
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Developing Benchmark for BANGLABOOK",
+        "text_level": 1,
+        "bbox": [
+            112,
+            662,
+            381,
+            693
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A series of baseline models with combinations of different lexical and semantic features are chosen to evaluate the BANGLABOOK dataset. An overview of the models, evaluation metrics, results, and analysis of the experimental results are provided in this section.",
+        "bbox": [
+            112,
+            699,
+            489,
+            794
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Baseline Models & Features",
+        "text_level": 1,
+        "bbox": [
+            112,
+            807,
+            379,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the lexical features, we extract bag-of-words (BoW), char  $n$ -grams (1-3), and word  $n$ -grams (1-3) from the reviews as these representations have performed well in different classification tasks (Islam et al., 2022). After extracting the features, they are vectorized using TF-IDF and count vec",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/801e495ea340c69d691c05085acf634158a1c49e822d0ae50a93be87c223f433.jpg",
+        "image_caption": [
+            "Figure 2: Sentiment Distribution of top 5 most popular genres. In clockwise order, वस्थकालीन अप्रति स (Contemporary Novel), वस्थकालीन अप्रति तूधोत्तिक तूधोत्तिक तूधोत्तिक तूधोत्तिक तूधोत्तिक (Islamic Ideals and Doctrines), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक तूधोत्तिक (Islamic Books: Self-Development), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक (Student Life Development), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक (Translated Books: Self-Development and Meditation)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            275,
+            873,
+            495
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "torizer and trained on a series of ML models such as Random Forest (Breiman, 2001), XGBoost (Chen and Guestrin, 2016), linear SVM (Cortes and Vapnik, 1995), Logistic Regression (le Cessie and van Houwelingen, 1992) and Multinomial Naive Bayes (John and Langley, 1995). We choose LSTM (Hochreiter and Schmidhuber, 1997) with GloVe (Pennington et al., 2014) embedding for its ability to understand context along with recent dependency. We also fine-tuned two available transformer-based models in Bangla: Bangla-BERT (base-uncased) (110M parameters) (Sarker, 2020) and Bangla-BERT (large) (2.5B parameters) (Bhattacharjee et al., 2022), due to the recent success of BERT (Devlin et al., 2019) in various downstream NLP tasks. We select F1-score and weighted average F1-score to evaluate the models because the dataset has an un",
+        "bbox": [
+            507,
+            629,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1239",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/94ccfbba47717f6cbdfacbfacdb6db65defabf5cc1984a315d1b94240c5530af.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Class</td><td>10 Most Frequent Words</td></tr><tr><td>Positive</td><td>बालेला (good), शन्तर (nice), अस्थान्यारुँ (extraordinary), दूपाकोँ (splendid), वर्णा (best), सहानुँ (facile), धनाकोर (beautiful), अश्थि (hope), वन्तर्यादित (gratefulness), अलामान्यारुँ, वल्तर्यादित (gratitude)</td></tr><tr><td>Neutral</td><td>बालेला (good), शन्तर (nice), अशरागोँ (bad), वन्तार्यादित (average), वल (fault), वर्णारेम (hope), सहानुँ (facile), धनाकोर (extraordinary), कम्ज (low), वन्तर्यादित (splendid)</td></tr><tr><td>Negative</td><td>बालेला (good), वल्तर (trash), अशरागोँ (bad), वल (fault), वर्णारेम (nice), वर्णारेम (hope), वति (waste), कम्ज (useless), वन्तर्यादित (disappointment), कम्ज (low)</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            223
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Most frequent word unigrams conveying the strongest sentiments of each class with English translation. The colors respectively denote Positive, Neutral and Negative sentiments.",
+        "bbox": [
+            112,
+            225,
+            882,
+            258
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e56d3518d85c08bfa6d3d061bdfe7945c590061f3d16e56b768f41c191a92c4f.jpg",
+        "image_caption": [
+            "Figure 3: Confusion Matrix for Bangla-BERT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            269,
+            411,
+            418
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "even class distribution. F1-score is the harmonic mean of precision and recall and it helps balance the metric across the imbalanced positive/negative samples (Sokolova et al., 2006). All our experiments are done using scikit-learn, pytorch, and transformers (Vaswani et al., 2017) and run on Google Colaboratory. The training, testing, and validation split of the entire dataset was 70-20-10 with previously unseen samples in the test and validation set.",
+        "bbox": [
+            112,
+            451,
+            489,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Results & Findings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            624,
+            310,
+            639
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 3 summarizes the experimental results for BANGLABOOK. Results show that BanglaBERT (large) outperforms all other models by a clear margin. Also, the combination of word/char 2-gram and word/char 3-gram perform exceptionally well with respective classifiers. Our hypothesis is that these two features result in a large number of unique word and character combinations, aiding the models' ability to generalize effectively across categories. Islam et al. (2022, 2021); Taher et al. (2018) concur on the same verdict by implying that the task predominantly relies on word units, with minimal dependence on subword level information and the nature of the Bangla language itself. Furthermore, Majumder et al. (2002) outlined the suitability of  $n$ -gram approaches in generating language profiles for Indo",
+        "bbox": [
+            112,
+            645,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "European languages in their work. The bag-of-words (BoW) feature is inept at classifying the corresponding categories because of its inability to capture critical contextual information and nuance (Zheng and Casari, 2018). Although word 1-gram does not outperform word 2-gram and word 3-gram, it does predict the 'Neutral' class well. Both the pre-trained Bangla-BERT models perform fairly consistently across all categories on the BANGLABOOK dataset, demonstrating the usefulness of contextual understanding and transfer learning in classification tasks even in low-resource languages like Bangla. The LSTM model with GloVe embedding recognizes the 'Negative' and 'Positive' classes very marginally and fails completely to identify the 'Neutral' category. It is also notable that, SVM with bigram and trigram achieves perfect scores in the 'Positive' class.",
+        "bbox": [
+            507,
+            269,
+            884,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To summarize, the utilization of pre-trained models (i.e. Bangla-BERT) that undergo training on extensive corpora, leading to exposure to extensive general language knowledge, has significantly contributed to their superior classification performance compared to other models and word embeddings. Additionally, models trained on handcrafted features also perform significantly well. It should be noted that Bangla pre-trained models are currently undergoing development, and further training on expansive corpora has the potential to enhance their ability to generalize and achieve even more impressive results.",
+        "bbox": [
+            507,
+            560,
+            882,
+            769
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Error Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            784,
+            672,
+            799
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the 'Positive' class, all the models produce excellent classification results. While some models perform reasonably well on the 'Negative' class, nearly all of the models perform poorly on the 'Neutral' class. The class imbalance of the dataset, as shown in Figure 1, is one obvious cause of this fluctuation in results. The confusion matrix for",
+        "bbox": [
+            507,
+            806,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1240",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Bangla-BERT on our dataset, presented in Figure3, reveals that most of the 'Negative' and 'Neutral' samples are misclassified as 'Positive' samples by our classifiers. To further analyze the misclassifications, we examine the W1 (word unigrams) of these three classes. We find 124,796 unique W1 for the 'Positive' class, 20,714 unique W1 for the 'Negative' class, and 19,096 unique W1 for the 'Neutral' class.  $77.57\\%$  of the W1 from the 'Neutral' class and  $79.83\\%$  of the W1 from the 'Negative' class are found in the 'Positive' class. Table 4 depicts the most frequent W1 conveying the strongest sentiments for each class. With only one distinct 'Neutral' W1 and even the 'Negative' class having multiple positive W1, the dominance of 'Positive' sentiment W1 over the other two classes is evident. This may have contributed to the lack of distinctive words in the 'Negative' and 'Neutral' classes, which inevitably prevented the feature-based models from generalizing.",
+        "bbox": [
+            110,
+            84,
+            492,
+            407
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Morphology and Negation Patterns of Bangla",
+        "text_level": 1,
+        "bbox": [
+            112,
+            417,
+            478,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Understanding the morphology and negation patterns of a language holds paramount importance in the realm of sentiment analysis because negation can alter the meaning of words and phrases, thereby affecting the overall sentiment conveyed by a text. We provide a concise yet insightful recapitulation of the topic in the case of Bangla accompanied by review samples from our dataset BANGLABOOK as the respective examples. From the linguistic typological standpoint, Bangla is categorized as a subject-object-verb (SOV) language because the subject, object, and verb generally adhere to said order in its sentential structure (Ramchand, 2004). The most common juxtaposition of polarity from positive to negative is the use of  $n_i$  ( $\\widehat{n}$ ) as a tensed negative. For example,",
+        "bbox": [
+            110,
+            458,
+            489,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Aaannn aonnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnaee",
+        "bbox": [
+            163,
+            718,
+            438,
+            777
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Another negational feature is expressed by placing na (nata) prior to the non-finite verb and after the finite verb in a sentence (although there are some exceptions). For example,",
+        "bbox": [
+            112,
+            785,
+            487,
+            850
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "AeBnBnBn BnBnBnBn BnBnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn",
+        "bbox": [
+            154,
+            851,
+            445,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The Bangla language consists of no negative adverbs or pronouns (Thompson, 2006). This is why the negative element responsible for the reversal of polarity transcends from the word-level to the sentence-level rendering the occurrences of almost all negations in Bangla manifest on the syntactic level (Thompson, 2006).",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the cases of double negatives, we see the involvement of lexical negation, a morphological feature that works with negative affixes (prefixes and suffixes) attached to a root word. The prefixes in Bangla have two different phonetic variations or allophones depending on whether the prefix precedes a vowel or a consonant. The same is true for prefixes that imbue a negative connotation to a root word, e.g.  $o$  (ə) and  $on$  (ən). For example,",
+        "bbox": [
+            507,
+            198,
+            884,
+            341
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Kekc 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e",
+        "bbox": [
+            551,
+            341,
+            840,
+            464
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Another negative prefix that precedes a root word to invert its polarity is  $nir$  (NIR). For example,",
+        "bbox": [
+            507,
+            474,
+            882,
+            505
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Leye 1",
+        "bbox": [
+            566,
+            507,
+            823,
+            551
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "On the contrary, the suffix  $hin$  (रिन) succeeds a root word to convert it to the corresponding negative form. For example,",
+        "bbox": [
+            507,
+            558,
+            882,
+            606
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Arnrkrn m  n knnnck gne  nna pae aen",
+        "bbox": [
+            546,
+            613,
+            858,
+            652
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The expression of negative sentiment is, therefore, very nuanced in the Bangla language as every occurrence of negative is intertwined with features like the tense, hierarchy of syntax, verb status, case-specific issues, and sequential arrangement of words (Thompson, 2006).",
+        "bbox": [
+            507,
+            658,
+            884,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            766,
+            640,
+            780
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This paper introduces BANGLABOOK, the largest Bangla book review dataset with 158,065 samples, each labeled with 1 of 3 user sentiments. We provide extensive statistical analysis and strong baselines facilitating the utility of the dataset. Given its massive size and fine-grained sentiment distribution, BANGLABOOK has the potential to alleviate the resource scarcity in Bangla language research.",
+        "bbox": [
+            507,
+            790,
+            884,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1241",
+        "bbox": [
+            482,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            250,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Many of the reviews that were gathered for constructing BANGLABOOK are discarded because they lack a corresponding rating. A manual annotation process would have yielded a much larger dataset, which was not feasible due to resource constraints. Moreover, one of the challenges for validating the dataset is the lack of statistical models and word-embeddings pre-trained on the Bangla language. Some pre-trained Bangla-BERT models, yet to be trained on extensive corpora, have only recently been proposed. Improving transformer-based models for Bangla can enhance sub-word level contextual understanding which will consequently help in more accurate identification of the sentiments in BANGLABOOK (Islam et al., 2022).",
+        "bbox": [
+            112,
+            109,
+            492,
+            367
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            84,
+            608,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mst Tuhin Akter, Manoara Begum, and Rashed Mustafa. 2021. Bengali sentiment analysis of e-commerce product reviews using k-nearest neighbors. In 2021 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD), pages 40-44. IEEE.",
+            "Sanjida Akter and Muhammad Tareq Aziz. 2016. Sentiment analysis on facebook group using lexicon based approach. In 2016 3rd International Conference on Electrical Engineering and Information Communication Technology (ICEEICT), pages 1-4. IEEE.",
+            "Shad Al Kaiser, Sudipta Mandal, Ashraful Kalam Abid, Ekhfa Hossain, Ferdous Bin Ali, and Intisar Tahmid Naheen. 2021. Social media opinion mining based on bangla public post of facebook. In 2021 24th International Conference on Computer and Information Technology (ICCIT), pages 1-6. IEEE.",
+            "Abhik Bhattacharjee, Tahmid Hasan, Wasi Ahmad, Kazi Samin Mubasshir, Md Saiful Islam, Anindya Iqbal, M. Sohel Rahman, and Rifat Shahriyar. 2022. BanglaBERT: Language model pretraining and benchmarks for low-resource language understanding evaluation in Bangla. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 1318-1327, Seattle, United States. Association for Computational Linguistics.",
+            "L. Breiman. 2001. Random forests. Machine Learning, 45:5-32.",
+            "Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.",
+            "Shaika Chowdhury and Wasifa Chowdhury. 2014. Performing sentiment analysis in bangla microblog posts. In 2014 International Conference on Informatics, Electronics & Vision (ICIEV), pages 1-6. IEEE.",
+            "Corinna Cortes and Vladimir Naumovich Vapnik. 1995. Support-vector networks. Machine Learning, 20:273-297.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805.",
+            "Rajib Chandra Dey and Orvila Sarker. 2019. Sentiment analysis on bengali text using lexicon based approach. In 2019 22nd International Conference on Computer and Information Technology (ICCIT), pages 1-5. IEEE.",
+            "Asif Hassan, Mohammad Rashedul Amin, Abul Kalam Al Azad, and Nabeel Mohammed. 2016. Sentiment analysis on bangla and romanized bangla text using"
+        ],
+        "bbox": [
+            510,
+            107,
+            885,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1242",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "deep recurrent models. In 2016 International Workshop on Computational Intelligence (IWCI), pages 51-56. IEEE.",
+            "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9:1735-1780.",
+            "MD Iqbal, Avishek Das, Omar Sharif, Mohammed Moshiul Hoque, and Iqbal H Sarker. 2022. Bemoc: A corpus for identifying emotion in bengali texts. SN Computer Science, 3(2):1-17.",
+            "Khondoker Ittehadul Islam, Md Saiful Islam, and Md Ruhul Amin. 2020. Sentiment analysis in bengali via transfer learning using multi-lingual bert. In 2020 23rd International Conference on Computer and Information Technology (ICCIIT), pages 1-5. IEEE.",
+            "Khondoker Ittehadul Islam, Sadipta Kar, Md Saiful Islam, and Mohammad Ruhul Amin. 2021. Sentnob: A dataset for analysing sentiment on noisy bangla texts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3265-3271.",
+            "Khondoker Ittehadul Islam, Tanvir Yuvraz, Md Saiful Islam, and Enamul Hassan. 2022. Emonoba: A dataset for analyzing fine-grained emotions on noisy bangla texts. 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, pages 128-134.",
+            "George H. John and Pat Langley. 1995. Estimating continuous distributions in bayesian classifiers. In _Conference on Uncertainty in Artificial Intelligence_.",
+            "Mst Eshita Khatun and Tapasy Rabeya. 2022. A machine learning approach for sentiment analysis of book reviews in bangla language. In 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), pages 1178-1182. IEEE.",
+            "Saskia le Cessie and J. C. van Houwelingen. 1992. Ridge estimators in logistic regression. Applied statistics, 41:191-201.",
+            "Tom MY Lin, Pin Luarn, and Yun Kuei Huang. 2005. Effect of internet book reviews on purchase intention: A focus group study. The Journal of Academic Librarianship, 31(5):461-468.",
+            "Shamsul Arafin Mahtab, Nazmul Islam, and Md Mahfuzur Rahaman. 2018. Sentiment analysis on bangladesh cricket with support vector machine. In 2018 international conference on Bangla speech and language processing (ICBSLP), pages 1-4. IEEE.",
+            "P Majumder, M Mitra, and BB Chaudhuri. 2002. Ngram: a language independent approach to ir and nlp. In International conference on universal knowledge and language, volume 2."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Muhammad Mahmudun Nabi, Md Tanzir Altaf, and Sabir Ismail. 2016. Detecting sentiment from bangla text using machine learning technique and feature analysis. International Journal of Computer Applications, 153(11):28-34.",
+            "Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing.",
+            "Md Rahib, Rumman Hussain Khan, Amzad Hussain Tamim, Mohammad Zawad Tahmeed, and Mohammad Jaber Hossain. 2022. Emotion detection based on bangladeshi peoples social media response on Covid-19. SN Computer Science, 3(2):1-6.",
+            "Fuad Rahman, Habibur Khan, Zakir Hossain, Mahfuza Begum, Sadia Mahanaz, Ashraful Islam, and Aminul Islam. 2019. An annotated bangla sentiment analysis corpus. In 2019 International Conference on Bangla Speech and Language Processing (ICBSLP), pages 1-5. IEEE.",
+            "Md Atikur Rahman and Emon Kumar Dey. 2018. Datasets for aspect-based sentiment analysis in bangla and its baseline evaluation. Data, 3(2):15.",
+            "Gillian Catriona Ramchand. 2004. Two types of negation in bengali. *Clause structure in South Asian languages*, pages 39-66.",
+            "Sagor Sarker. 2020. Banglabert: Bengali mask language model for bengali language understanding.",
+            "Salim Sazzed. 2020a. Cross-lingual sentiment classification in low-resource bengali language. In Proceedings of the sixth workshop on noisy user-generated text (W-NUT 2020), pages 50–60.",
+            "Salim Sazzed. 2020b. Development of sentiment lexicon in bengali utilizing corpus and cross-lingual resources. In 2020 IEEE 21st International conference on information reuse and integration for data science (IRI), pages 237-244. IEEE.",
+            "Salim Sazzed. 2021. Bengsentilex and bengswear-lex: creating lexicons for sentiment analysis and profanity detection in low-resource bengali language. PeerJ Computer Science, 7:e681.",
+            "Marina Sokolova, Nathalie Japkowicz, and S. Szpakowicz. 2006. Beyond accuracy, f-score and roc: A family of discriminant measures for performance evaluation. In Australian Conference on Artificial Intelligence.",
+            "Alan T Sorensen and Scott J Rasmussen. 2004. Is any publicity good publicity? a note on the impact of book reviews. NBER Working paper, Stanford University.",
+            "Nusrath Tabassum and Muhammad Ibrahim Khan. 2019. Design an empirical framework for sentiment analysis from bangla text using machine learning. In 2019 International Conference on Electrical, Computer and Communication Engineering (ECCE), pages 1-5. IEEE."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1243",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "SM Abu Taher, Kazi Afsana Akhter, and KM Azharul Hasan. 2018. N-gram based sentiment mining for bangla text using support vector machine. In 2018 international conference on Bangla speech and language processing (ICBSLP), pages 1-5. IEEE.",
+            "Hanne-Ruth Thompson. 2006. Negation patterns in bengali. Bulletin of the School of Oriental and African Studies, 69(2):243-265.",
+            "Rashedul Amin Tuhin, Bechitra Kumar Paul, Faria Nawrine, Mahbuba Akter, and Amit Kumar Das. 2019. An automated system of sentiment analysis from bangla text using supervised learning techniques. In 2019 IEEE 4th International Conference on Computer and Communication Systems (ICCCS), pages 360-364. IEEE.",
+            "Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. ArXiv, abs/1706.03762.",
+            "Anning Wang, Qiang Zhang, Shuangyao Zhao, Xiaonong Lu, and Zhanglin Peng. 2020. A review-driven customer preference measurement model for product improvement: sentiment-based importance performance analysis. Information Systems and e-Business Management, 18:61-88.",
+            "Alice Zheng and Amanda Casari. 2018. Feature engineering for machine learning: principles and techniques for data scientists. \"O'Reilly Media, Inc.\""
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            504
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "bbox": [
+            115,
+            517,
+            236,
+            533
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1244",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ce284ac3f938bb54ce76e1b58041e5f88ea2dac95ca89db7651aeb616a5cfbb1.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 5: Comparison of notable Bangla Sentiment Analysis datasets sorted in ascending order of size. The abbreviations respectively denote, Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), Naive Bayes (NB), Logistic Regression (LR), K-Nearest Neighbors (KNN), Long Short-term Memory (LSTM), Convolutional Neural Network (CNN), Extreme Gradient Boost (XGB), Multinomial Naive Bayes (MNB), Radial Basis Function (RBF), Extreme Random Tree (ET), Recurrent Neural Network (RNN), Bidirectional Encoder Representations from Transformers (BERT), Gated Recurrent Unit (GRU), Bi-LSTM (Bidirectional LSTM). All the publicly available datasets are hyperlinked. Open† denotes the redaction of the link for anonymity."
+        ],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"2\">Sentiment Classification</td><td colspan=\"2\">Sentiment Distribution</td><td>Total # of Samples</td><td>Availability</td><td>Type of Content</td><td>Source(s)</td><td>Baseline Models</td></tr><tr><td>(Tabassum and Khan, 2019)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>1,050</td><td>Closed</td><td>Posts, comments</td><td>Facebook, Twitter</td><td>RF</td></tr><tr><td>(Chowdhury and Chowdhury, 2014)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>1,300</td><td>Closed</td><td>Posts, comments</td><td>Twitter</td><td>-</td></tr><tr><td>(Nabi et al., 2016)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>1,500</td><td>Closed</td><td>Posts, comments</td><td>Social Media</td><td>-</td></tr><tr><td>(Mahtab et al., 2018)</td><td colspan=\"2\">Praise Criticism Sadness</td><td colspan=\"2\">513 604 484</td><td>1,601</td><td>Closed</td><td>Comments</td><td>Prothom Alo Online News Portal</td><td>SVM, DT, NB</td></tr><tr><td>(Akter and Aziz, 2016)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">- -</td><td>3,600</td><td>Closed</td><td>Posts, comments</td><td>Facebook</td><td>NB</td></tr><tr><td>(Rahman and Kumar Dey, 2018)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">- -</td><td>4,700</td><td>Open</td><td>Comments</td><td>Facebook pages: BBC Bangla, Prothom Alo</td><td>SVM, LR, KNN, DT, LSTM NB, CNN</td></tr><tr><td>(Dey and Sarker, 2019)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">2,600 2,600</td><td>5,200</td><td>Closed</td><td>Comments, reviews</td><td>Facebook, Twitter, YouTube, News Portals</td><td>DT, NB, SVM</td></tr><tr><td>(Khatun and Rabeya, 2022)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>5,500</td><td>Closed</td><td>Comments, reviews</td><td>Social Media</td><td>-</td></tr><tr><td>BEMOC (Iqbal et al., 2022)</td><td colspan=\"2\">AngerFearSadJoyDisgust</td><td colspan=\"2\">- -</td><td>7,000</td><td>Open</td><td>Posts, comments</td><td>Facebook, YouTube, Online blogs, Bangla story books, novels, newspapers, discourse</td><td>-</td></tr><tr><td>(Tuhin et al., 2019)</td><td colspan=\"2\">Happy Tender Excited Sad Angry Scared</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>(Akter et al., 2021)</td><td colspan=\"2\">Positive Negative Neutral</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td>7,905</td><td>Closed</td><td>Product reviews</td><td>Daraz</td><td>RF, LR, SVM, KNN, XGB</td><td></td></tr><tr><td>(Hassan et al., 2016)</td><td colspan=\"2\">Positive Negative Ambiguous</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>(Rahib et al., 2022)</td><td colspan=\"2\">Insightful Curious Gratitude</td><td colspan=\"2\">3,800 3,549 3,232</td><td>10,581</td><td>Closed</td><td>Comments</td><td>Social Media</td><td>SVM, RF, CNN, LSTM</td></tr><tr><td rowspan=\"2\">(Al Kaiser et al., 2021)</td><td rowspan=\"2\">Positive Negative Neutral</td><td>Wishful Thinking Appreciation</td><td>967 942</td><td>1,909</td><td rowspan=\"2\">11,006</td><td rowspan=\"2\">Closed</td><td rowspan=\"2\">Comments</td><td rowspan=\"2\">Facebook</td><td rowspan=\"2\">LR, DT, RF, MNB, KNN, Linear SVM, RBF SVM, XGB</td></tr><tr><td>Gender-based hate Religious hate Political hate Personal hate Sarcasm</td><td>525 731 572 1,995 1,414 3,860</td><td>5,237 3,860</td></tr><tr><td>(Sazzed, 2020a)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">8,500 3,307</td><td>11,807</td><td>Open</td><td>Comments</td><td>YouTube</td><td>SVM, ET, RF, LR, VADER, TextBlob</td></tr><tr><td>(Sazzed, 2020b)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>12,000</td><td>Closed</td><td>Comments</td><td>YouTube</td><td>-</td></tr><tr><td rowspan=\"2\">SENTNoB (Islam et al., 2021)</td><td rowspan=\"2\" colspan=\"2\">Positive Negative</td><td rowspan=\"2\" colspan=\"2\">6,410 5,709 3,609</td><td rowspan=\"2\">15,728</td><td rowspan=\"2\">Open</td><td rowspan=\"2\">Comments</td><td rowspan=\"2\">Prothom Alo Online Newspaper, YouTube</td><td rowspan=\"2\">RNN</td></tr><tr></tr><tr><td>(Islam et al., 2020)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">4,769 8,351 4,732</td><td>17,852</td><td>Open</td><td>Comments</td><td>Prothom Alo Online Newspaper</td><td>CNN, LSTM, BERT, GRU, fastText</td></tr><tr><td>EMONOBA (Islam et al., 2022)</td><td colspan=\"2\">Love Joy Surprise Anger Sadness Fear</td><td colspan=\"2\">4,202 9,249 939 3,905 5,109 307</td><td>20,468</td><td>Open</td><td>Comments</td><td>YouTube, Facebook, Twitter, Prothom Alo</td><td>Bi-LSTM, fastText, Bangla-BERT-base</td></tr><tr><td>BANGLABOOK (ours)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">141,587 9,674 6,804</td><td>158,065</td><td>\\( Open^† \\)</td><td>Book reviews</td><td>Rokomari, Wafilife</td><td>RF, LSTM, LR, GRU MNB, SVM, XGB Bangla-BERT</td></tr></table>",
+        "bbox": [
+            114,
+            142,
+            931,
+            758
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1245",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            143
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            143,
+            231,
+            159
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            127,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            187,
+            349,
+            200
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            212,
+            695,
+            229
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            230,
+            231,
+            244
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            256,
+            668,
+            272
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            213,
+            336
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            127,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            363,
+            349,
+            379
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            127,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            407,
+            349,
+            422
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            127,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            497,
+            349,
+            513
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            127,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            573,
+            349,
+            588
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            127,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            632,
+            349,
+            646
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            129,
+            657,
+            880,
+            739
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            740,
+            231,
+            753
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            785,
+            213,
+            801
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            813,
+            880,
+            845
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            846,
+            231,
+            860
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1246",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank."
+        ],
+        "bbox": [
+            129,
+            84,
+            880,
+            282
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.",
+        "bbox": [
+            112,
+            293,
+            877,
+            330
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank."
+        ],
+        "bbox": [
+            127,
+            340,
+            880,
+            640
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1247",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/BanglaBook_ A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews/28378eaf-0a48-458e-a698-291e5038e7bf_model.json

@@ -0,0 +1,2004 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.14,
+                0.083,
+                0.861,
+                0.121
+            ],
+            "angle": 0,
+            "content": "BANGLABOOK: A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.241,
+                0.131,
+                0.764,
+                0.165
+            ],
+            "angle": 0,
+            "content": "Mohsinul Kabir*, Obayed Bin Mahfuz*, Syed Rifat Raiyan*, Hasan Mahmud, Md Kamrul Hasan"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.365,
+                0.166,
+                0.639,
+                0.182
+            ],
+            "angle": 0,
+            "content": "Systems and Software Lab (SSL)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.297,
+                0.183,
+                0.708,
+                0.199
+            ],
+            "angle": 0,
+            "content": "Department of Computer Science and Engineering"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.282,
+                0.199,
+                0.721,
+                0.215
+            ],
+            "angle": 0,
+            "content": "Islamic University of Technology, Dhaka, Bangladesh"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.185,
+                0.216,
+                0.821,
+                0.232
+            ],
+            "angle": 0,
+            "content": "{mohsinulkabir, siam, rifatraiyan, hasan, hasank}@iut-dhaka.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.281,
+                0.461,
+                0.678
+            ],
+            "angle": 0,
+            "content": "The analysis of consumer sentiment, as expressed through reviews, can provide a wealth of insight regarding the quality of a product. While the study of sentiment analysis has been widely explored in many popular languages, relatively less attention has been given to the Bangla language, mostly due to a lack of relevant data and cross-domain adaptability. To address this limitation, we present BANGLABOOK, a large-scale dataset of Bangla book reviews consisting of 158,065 samples classified into three broad categories: positive, negative, and neutral. We provide a detailed statistical analysis of the dataset and employ a range of machine learning models to establish baselines including SVM, LSTM, and Bangla-BERT. Our findings demonstrate a substantial performance advantage of pretrained models over models that rely on manually crafted features, emphasizing the necessity for additional training resources in this domain. Additionally, we conduct an in-depth error analysis by examining sentiment unigrams, which may provide insight into common classification errors in under-resourced languages like Bangla. Our codes and data are publicly available at https://github.com/mohsinulkabir14/BanglaBook."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.69,
+                0.26,
+                0.705
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.715,
+                0.49,
+                0.86
+            ],
+            "angle": 0,
+            "content": "The resources publicly available for scholarly investigation in the realm of Sentiment Analysis (SA) for the Bangla language are scarce and limited in quantity (Khatun and Rabeya, 2022; Sazzed, 2021; Rahman et al., 2019) despite its literary gravitas as the \\(6^{\\text{th}}\\) most spoken language\\(^{1}\\) in the world with approximately 200 million speakers. In the existing literature on Bangla Text SA, as shown in Table 5, the largest dataset consists"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.253,
+                0.886,
+                0.768
+            ],
+            "angle": 0,
+            "content": "of 20,468 samples (Islam et al., 2022) while the smallest has a mere 1,050 samples (Tabassum and Khan, 2019). Besides these, Islam et al. (2020) created a dataset consisting of 17,852 samples and Islam et al. (2021) utilized a dataset of 15,728 samples. All other datasets apart from these either have \\(< 15,000\\) samples or are publicly unavailable. Another limitation of the existing research works in Bangla Text SA is the deficiency of datasets having product-specific review samples. Most of the available Bangla SA datasets are focused on user-generated textual content from cyberspace. The insights derived from these may not accurately represent sentiment in the context of product reviews, thus hindering their usefulness for businesses. The tonal and linguistic analysis of reviews from product-specific datasets can aid businesses to gain valuable insights into customer attitudes, preferences, and experiences which can then be leveraged to improve products and services, design targeted marketing campaigns, and make more informed business decisions. In this paper, we introduce a large-scale dataset, BANGLABOOK, consisting of 158,065 samples of book reviews collected from online bookshops written in the Bangla language. This is the largest dataset for Bangla sentiment analysis to the best of our knowledge. We perform an analysis of the dataset's statistical characteristics, employ various ML techniques to establish a performance benchmark for validating the dataset, and also conduct a thorough evaluation of the classification errors."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.78,
+                0.729,
+                0.794
+            ],
+            "angle": 0,
+            "content": "2 Dataset Construction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.805,
+                0.885,
+                0.885
+            ],
+            "angle": 0,
+            "content": "In order to create this dataset, we collect a total of 204,659 book reviews from two online bookshops (Rokomari² and Wafilife³) using a web scraper developed with several Python libraries, including BeautifulSoup, Selenium, Pandas, Openpyxl,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.53,
+                0.891,
+                0.732,
+                0.905
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://www.rokomari.com/"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.53,
+                0.905,
+                0.732,
+                0.918
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>https://www.wafilife.com/"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.53,
+                0.891,
+                0.732,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.868,
+                0.488,
+                0.894
+            ],
+            "angle": 0,
+            "content": "*These authors contributed equally to this work. Author names are in alphabetic order."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.894,
+                0.434,
+                0.906
+            ],
+            "angle": 0,
+            "content": "1https://en.wikipedia.org/wiki/List_of_"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.906,
+                0.4,
+                0.919
+            ],
+            "angle": 0,
+            "content": "languages_by_total_number_of-speakers"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.868,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1237"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.229,
+                0.946,
+                0.769,
+                0.959
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 1237-1247"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.297,
+                0.96,
+                0.702,
+                0.973
+            ],
+            "angle": 0,
+            "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.274,
+                0.082,
+                0.725,
+                0.235
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Source</td><td>Language</td><td>Annotated</td><td>Unannotated</td><td>Total</td></tr><tr><td rowspan=\"4\">Rokomari</td><td>Bangla</td><td>84,672</td><td>24,806</td><td>109,478</td></tr><tr><td>Bangla† + English</td><td>56,485</td><td>15,408</td><td>71,893</td></tr><tr><td>Bangla + Bangla† + English</td><td>13,144</td><td>4,114</td><td>17,258</td></tr><tr><td>Total</td><td>154,301</td><td>44,328</td><td>198,629</td></tr><tr><td rowspan=\"4\">Wafilife</td><td>Bangla</td><td>4,699</td><td>59</td><td>4,758</td></tr><tr><td>Bangla† + English</td><td>370</td><td>3</td><td>373</td></tr><tr><td>Bangla + Bangla† + English</td><td>896</td><td>3</td><td>899</td></tr><tr><td>Total</td><td>5,965</td><td>65</td><td>6,030</td></tr><tr><td rowspan=\"4\"></td><td>Bangla</td><td>89,371</td><td>24,865</td><td>114,237</td></tr><tr><td>Bangla† + English</td><td>56,855</td><td>15,411</td><td>72,266</td></tr><tr><td>Bangla + Bangla† + English</td><td>14,040</td><td>4,117</td><td>18,157</td></tr><tr><td>Total</td><td>160,266</td><td>44,393</td><td>204,659</td></tr><tr><td></td><td>Untranslated Data (Removed)</td><td>2,201</td><td></td><td></td></tr><tr><td></td><td>Final Dataset Size</td><td>158,065</td><td></td><td></td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.212,
+                0.238,
+                0.782,
+                0.253
+            ],
+            "angle": 0,
+            "content": "Table 1: Summary statistics of our dataset. Bangla† denotes Romanized Bangla text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.267,
+                0.486,
+                0.297
+            ],
+            "angle": 0,
+            "content": "and Webdriver, to collect and process the raw data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.3,
+                0.49,
+                0.428
+            ],
+            "angle": 0,
+            "content": "For the data collection and preparation process of the BANGLABOOK dataset, we first compile a list of URLs for authors from online bookstores. From there, we procure URLs for the books. We meticulously scrape information such as book titles, author names, book categories, review texts, reviewer names, review dates, and ratings by utilizing these book URLs."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.162,
+                0.44,
+                0.443,
+                0.54
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"5\">General Properties</td><td># of Reviews</td><td>158,065</td></tr><tr><td># of Books</td><td>30,253</td></tr><tr><td># of Reviewers</td><td>44,616</td></tr><tr><td># of Categories</td><td>1,573</td></tr><tr><td>Total Review Words</td><td>44,429,201</td></tr><tr><td rowspan=\"3\">Single Review</td><td>max. Word length</td><td>765</td></tr><tr><td>min. Word length</td><td>1</td></tr><tr><td>avg. Word length</td><td>281.08</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.143,
+                0.543,
+                0.457,
+                0.556
+            ],
+            "angle": 0,
+            "content": "Table 2: General overview of BANGLABOOK."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.576,
+                0.345,
+                0.591
+            ],
+            "angle": 0,
+            "content": "2.1 Labeling & Translation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.598,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "If a review does not have a rating, we deem it unannotated. Reviews with a rating of 1 or 2 are classified as negative, a rating of 3 is considered neutral, and a rating of 4 or 5 is classified as positive. Two manual experiments are carried out to validate the use of ratings as a measure of sentiment in product reviews. In the first experiment, around \\(10\\%\\) of the reviews are randomly selected and annotated manually. The annotated labels are cross-checked with the original labels, resulting in a \\(96.7\\%\\) accuracy in the corresponding labels. In addition, we consult the work of Wang et al. (2020) that explored the issue of incongruous sentiment expressions with regard to ratings. Specifically, the study scrutinized two categories of reviews: high ratings lacking a positive sentiment, and low ratings lacking a negative sentiment. We perform an analysis to identify such inconsistencies within our dataset and discovered that only a minuscule \\(3.41\\%\\) of the samples exhibited this pattern. This figure is rela"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.267,
+                0.884,
+                0.299
+            ],
+            "angle": 0,
+            "content": "tively insignificant when considering the substantially large scale of our dataset."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.311,
+                0.679,
+                0.432
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.518,
+                0.439,
+                0.678,
+                0.451
+            ],
+            "angle": 0,
+            "content": "(a) Sentiment Distribution"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.682,
+                0.319,
+                0.876,
+                0.433
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.708,
+                0.439,
+                0.848,
+                0.451
+            ],
+            "angle": 0,
+            "content": "(b) Rating Distribution"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.533,
+                0.457,
+                0.857,
+                0.471
+            ],
+            "angle": 0,
+            "content": "Figure 1: Class Distribution of BANGLABOOK."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.476,
+                0.885,
+                0.782
+            ],
+            "angle": 0,
+            "content": "After discarding the unannotated reviews, we curate a final dataset of 158,065 annotated reviews. Of these, 89,371 are written entirely in Bangla. The remaining 68,694 reviews were written in Romanized Bangla, English, or a mix of languages. They are translated into Bangla with Google Translator and a custom Python program using the googletrans library. The translations are subsequently subjected to manual review and scrutiny to confirm their accuracy. The majority of inaccurate translations primarily comprise spelling errors and instances where English words remain untranslated within samples containing a combination of Bangla and English text. The meticulous evaluation process of untranslated samples involves a thorough assessment by post-graduate native Bangla speakers, who critically compare the translated text against the original untranslated text to ascertain the correctness of the translation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.796,
+                0.709,
+                0.813
+            ],
+            "angle": 0,
+            "content": "3 Statistical Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Tables 1 and 2 provide an overview of the statistical properties of the BANGLABOOK dataset. The sentiment doughnut chart in Figure-1a illustrates the proportion of positive, neutral, and negative reviews, while the rating doughnut chart in Figure-1b displays the percentage of reviews that corre"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1238"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.131,
+                0.082,
+                0.871,
+                0.232
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>Negative</td><td>Neutral</td><td>Positive</td><td>Weighted Avg.</td></tr><tr><td>Random Forest (word 2-gram + word 3-gram)</td><td>0.56</td><td>0.34</td><td>0.96</td><td>0.9106</td></tr><tr><td>SVM (word 2-gram + word 3-gram)</td><td>0.40</td><td>0.15</td><td>1.00</td><td>0.9053</td></tr><tr><td>Random Forest (word 1-gram)</td><td>0.48</td><td>0.35</td><td>0.96</td><td>0.9043</td></tr><tr><td>Logistic Regression (char 2-gram + char 3-gram)</td><td>0.55</td><td>0.13</td><td>0.96</td><td>0.8978</td></tr><tr><td>Bangla-BERT (base-uncased)</td><td>0.60</td><td>0.22</td><td>0.96</td><td>0.9064</td></tr><tr><td>Logistic Regression (word 2-gram + word 3-gram)</td><td>0.53</td><td>0.13</td><td>0.96</td><td>0.8964</td></tr><tr><td>Bangla-BERT (large)</td><td>0.72</td><td>0.40</td><td>0.97</td><td>0.9331</td></tr><tr><td>XGBoost (char 2-gram + char 3-gram)</td><td>0.31</td><td>0.02</td><td>0.95</td><td>0.8723</td></tr><tr><td>Multinomial NB (word 2-gram + word 3-gram)</td><td>0.23</td><td>0.03</td><td>0.95</td><td>0.8663</td></tr><tr><td>LSTM (GloVe)</td><td>0.11</td><td>0.00</td><td>0.10</td><td>0.0991</td></tr><tr><td>XGBoost (word 2-gram + word 3-gram)</td><td>0.23</td><td>0.01</td><td>0.95</td><td>0.8651</td></tr><tr><td>Multinomial NB (BoW)</td><td>0.18</td><td>0.05</td><td>0.94</td><td>0.8564</td></tr><tr><td>SVM (word 1-gram)</td><td>0.08</td><td>0.04</td><td>0.94</td><td>0.8519</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.236,
+                0.882,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Table 3: Catergory-wise Binary Task F1-score and Weighted Average F1-score of each method on BANGLABOOK."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.265,
+                0.415,
+                0.279
+            ],
+            "angle": 0,
+            "content": "spond to each rating on a scale of 1 to 5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.28,
+                0.49,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Upon analyzing the sentiment chart, it appears that the majority of the reviews \\((124,084 + 17,503 = 141,587\\) samples) are positive, with a significant portion also being negative \\((2,728 + 6,946 = 9,674\\) samples). A relatively small fraction of the reviews are neutral (6,804 samples). This suggests that overall, the books have been well received by the readers, with the majority expressing favorable opinions. The distribution of the dataset is representative of real-world scenarios and it tessellates well with previous content analysis works on book reviews (Lin et al., 2005; Sorensen and Rasmussen, 2004). In Figure-2, we can visualize an illustration of the sentiment distribution among the 5 most frequently reviewed categories of books. We can gain some salient insights from the popularity of these genres. Contemporary novels are bestsellers as they reflect current events, social issues, and trends, making them relatable and thought-provoking for the readers while self-help and religious books provide guidance, inspiration, and a sense of purpose, catering to individuals' quest for personal growth and spiritual fulfillment."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.663,
+                0.383,
+                0.694
+            ],
+            "angle": 0,
+            "content": "4 Developing Benchmark for BANGLABOOK"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.7,
+                0.49,
+                0.795
+            ],
+            "angle": 0,
+            "content": "A series of baseline models with combinations of different lexical and semantic features are chosen to evaluate the BANGLABOOK dataset. An overview of the models, evaluation metrics, results, and analysis of the experimental results are provided in this section."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.808,
+                0.38,
+                0.822
+            ],
+            "angle": 0,
+            "content": "4.1 Baseline Models & Features"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.824,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "For the lexical features, we extract bag-of-words (BoW), char \\( n \\)-grams (1-3), and word \\( n \\)-grams (1-3) from the reviews as these representations have performed well in different classification tasks (Islam et al., 2022). After extracting the features, they are vectorized using TF-IDF and count vec"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.526,
+                0.277,
+                0.875,
+                0.497
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.51,
+                0.885,
+                0.61
+            ],
+            "angle": 0,
+            "content": "Figure 2: Sentiment Distribution of top 5 most popular genres. In clockwise order, वस्थकालीन अप्रति स (Contemporary Novel), वस्थकालीन अप्रति तूधोत्तिक तूधोत्तिक तूधोत्तिक तूधोत्तिक तूधोत्तिक (Islamic Ideals and Doctrines), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक तूधोत्तिक (Islamic Books: Self-Development), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक (Student Life Development), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक (Translated Books: Self-Development and Meditation)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.63,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "torizer and trained on a series of ML models such as Random Forest (Breiman, 2001), XGBoost (Chen and Guestrin, 2016), linear SVM (Cortes and Vapnik, 1995), Logistic Regression (le Cessie and van Houwelingen, 1992) and Multinomial Naive Bayes (John and Langley, 1995). We choose LSTM (Hochreiter and Schmidhuber, 1997) with GloVe (Pennington et al., 2014) embedding for its ability to understand context along with recent dependency. We also fine-tuned two available transformer-based models in Bangla: Bangla-BERT (base-uncased) (110M parameters) (Sarker, 2020) and Bangla-BERT (large) (2.5B parameters) (Bhattacharjee et al., 2022), due to the recent success of BERT (Devlin et al., 2019) in various downstream NLP tasks. We select F1-score and weighted average F1-score to evaluate the models because the dataset has an un"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1239"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.081,
+                0.885,
+                0.224
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Class</td><td>10 Most Frequent Words</td></tr><tr><td>Positive</td><td>बालेला (good), शन्तर (nice), अस्थान्यारुँ (extraordinary), दूपाकोँ (splendid), वर्णा (best), सहानुँ (facile), धनाकोर (beautiful), अश्थि (hope), वन्तर्यादित (gratefulness), अलामान्यारुँ, वल्तर्यादित (gratitude)</td></tr><tr><td>Neutral</td><td>बालेला (good), शन्तर (nice), अशरागोँ (bad), वन्तार्यादित (average), वल (fault), वर्णारेम (hope), सहानुँ (facile), धनाकोर (extraordinary), कम्ज (low), वन्तर्यादित (splendid)</td></tr><tr><td>Negative</td><td>बालेला (good), वल्तर (trash), अशरागोँ (bad), वल (fault), वर्णारेम (nice), वर्णारेम (hope), वति (waste), कम्ज (useless), वन्तर्यादित (disappointment), कम्ज (low)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.226,
+                0.884,
+                0.259
+            ],
+            "angle": 0,
+            "content": "Table 4: Most frequent word unigrams conveying the strongest sentiments of each class with English translation. The colors respectively denote Positive, Neutral and Negative sentiments."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.195,
+                0.27,
+                0.413,
+                0.419
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.129,
+                0.422,
+                0.473,
+                0.437
+            ],
+            "angle": 0,
+            "content": "Figure 3: Confusion Matrix for Bangla-BERT"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.452,
+                0.49,
+                0.611
+            ],
+            "angle": 0,
+            "content": "even class distribution. F1-score is the harmonic mean of precision and recall and it helps balance the metric across the imbalanced positive/negative samples (Sokolova et al., 2006). All our experiments are done using scikit-learn, pytorch, and transformers (Vaswani et al., 2017) and run on Google Colaboratory. The training, testing, and validation split of the entire dataset was 70-20-10 with previously unseen samples in the test and validation set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.625,
+                0.312,
+                0.64
+            ],
+            "angle": 0,
+            "content": "4.2 Results & Findings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.646,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Table 3 summarizes the experimental results for BANGLABOOK. Results show that BanglaBERT (large) outperforms all other models by a clear margin. Also, the combination of word/char 2-gram and word/char 3-gram perform exceptionally well with respective classifiers. Our hypothesis is that these two features result in a large number of unique word and character combinations, aiding the models' ability to generalize effectively across categories. Islam et al. (2022, 2021); Taher et al. (2018) concur on the same verdict by implying that the task predominantly relies on word units, with minimal dependence on subword level information and the nature of the Bangla language itself. Furthermore, Majumder et al. (2002) outlined the suitability of \\( n \\)-gram approaches in generating language profiles for Indo"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.27,
+                0.885,
+                0.559
+            ],
+            "angle": 0,
+            "content": "European languages in their work. The bag-of-words (BoW) feature is inept at classifying the corresponding categories because of its inability to capture critical contextual information and nuance (Zheng and Casari, 2018). Although word 1-gram does not outperform word 2-gram and word 3-gram, it does predict the 'Neutral' class well. Both the pre-trained Bangla-BERT models perform fairly consistently across all categories on the BANGLABOOK dataset, demonstrating the usefulness of contextual understanding and transfer learning in classification tasks even in low-resource languages like Bangla. The LSTM model with GloVe embedding recognizes the 'Negative' and 'Positive' classes very marginally and fails completely to identify the 'Neutral' category. It is also notable that, SVM with bigram and trigram achieves perfect scores in the 'Positive' class."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.561,
+                0.884,
+                0.77
+            ],
+            "angle": 0,
+            "content": "To summarize, the utilization of pre-trained models (i.e. Bangla-BERT) that undergo training on extensive corpora, leading to exposure to extensive general language knowledge, has significantly contributed to their superior classification performance compared to other models and word embeddings. Additionally, models trained on handcrafted features also perform significantly well. It should be noted that Bangla pre-trained models are currently undergoing development, and further training on expansive corpora has the potential to enhance their ability to generalize and achieve even more impressive results."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.785,
+                0.673,
+                0.8
+            ],
+            "angle": 0,
+            "content": "4.3 Error Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In the 'Positive' class, all the models produce excellent classification results. While some models perform reasonably well on the 'Negative' class, nearly all of the models perform poorly on the 'Neutral' class. The class imbalance of the dataset, as shown in Figure 1, is one obvious cause of this fluctuation in results. The confusion matrix for"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1240"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.085,
+                0.493,
+                0.409
+            ],
+            "angle": 0,
+            "content": "Bangla-BERT on our dataset, presented in Figure3, reveals that most of the 'Negative' and 'Neutral' samples are misclassified as 'Positive' samples by our classifiers. To further analyze the misclassifications, we examine the W1 (word unigrams) of these three classes. We find 124,796 unique W1 for the 'Positive' class, 20,714 unique W1 for the 'Negative' class, and 19,096 unique W1 for the 'Neutral' class. \\(77.57\\%\\) of the W1 from the 'Neutral' class and \\(79.83\\%\\) of the W1 from the 'Negative' class are found in the 'Positive' class. Table 4 depicts the most frequent W1 conveying the strongest sentiments for each class. With only one distinct 'Neutral' W1 and even the 'Negative' class having multiple positive W1, the dominance of 'Positive' sentiment W1 over the other two classes is evident. This may have contributed to the lack of distinctive words in the 'Negative' and 'Neutral' classes, which inevitably prevented the feature-based models from generalizing."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.418,
+                0.479,
+                0.451
+            ],
+            "angle": 0,
+            "content": "5 Morphology and Negation Patterns of Bangla"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.459,
+                0.49,
+                0.717
+            ],
+            "angle": 0,
+            "content": "Understanding the morphology and negation patterns of a language holds paramount importance in the realm of sentiment analysis because negation can alter the meaning of words and phrases, thereby affecting the overall sentiment conveyed by a text. We provide a concise yet insightful recapitulation of the topic in the case of Bangla accompanied by review samples from our dataset BANGLABOOK as the respective examples. From the linguistic typological standpoint, Bangla is categorized as a subject-object-verb (SOV) language because the subject, object, and verb generally adhere to said order in its sentential structure (Ramchand, 2004). The most common juxtaposition of polarity from positive to negative is the use of \\( n_i \\) (\\( \\widehat{n} \\)) as a tensed negative. For example,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.164,
+                0.719,
+                0.44,
+                0.778
+            ],
+            "angle": 0,
+            "content": "Aaannn aonnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnaee"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.787,
+                0.489,
+                0.851
+            ],
+            "angle": 0,
+            "content": "Another negational feature is expressed by placing na (nata) prior to the non-finite verb and after the finite verb in a sentence (although there are some exceptions). For example,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.156,
+                0.852,
+                0.446,
+                0.918
+            ],
+            "angle": 0,
+            "content": "AeBnBnBn BnBnBnBn BnBnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.198
+            ],
+            "angle": 0,
+            "content": "The Bangla language consists of no negative adverbs or pronouns (Thompson, 2006). This is why the negative element responsible for the reversal of polarity transcends from the word-level to the sentence-level rendering the occurrences of almost all negations in Bangla manifest on the syntactic level (Thompson, 2006)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.199,
+                0.885,
+                0.342
+            ],
+            "angle": 0,
+            "content": "In the cases of double negatives, we see the involvement of lexical negation, a morphological feature that works with negative affixes (prefixes and suffixes) attached to a root word. The prefixes in Bangla have two different phonetic variations or allophones depending on whether the prefix precedes a vowel or a consonant. The same is true for prefixes that imbue a negative connotation to a root word, e.g. \\(o\\) (ə) and \\(on\\) (ən). For example,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.552,
+                0.342,
+                0.842,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Kekc 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.475,
+                0.883,
+                0.506
+            ],
+            "angle": 0,
+            "content": "Another negative prefix that precedes a root word to invert its polarity is \\(nir\\) (NIR). For example,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.508,
+                0.825,
+                0.552
+            ],
+            "angle": 0,
+            "content": "Leye 1"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.56,
+                0.884,
+                0.607
+            ],
+            "angle": 0,
+            "content": "On the contrary, the suffix \\(hin\\) (रिन) succeeds a root word to convert it to the corresponding negative form. For example,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.547,
+                0.614,
+                0.86,
+                0.653
+            ],
+            "angle": 0,
+            "content": "Arnrkrn m  n knnnck gne  nna pae aen"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.659,
+                0.885,
+                0.755
+            ],
+            "angle": 0,
+            "content": "The expression of negative sentiment is, therefore, very nuanced in the Bangla language as every occurrence of negative is intertwined with features like the tense, hierarchy of syntax, verb status, case-specific issues, and sequential arrangement of words (Thompson, 2006)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.767,
+                0.642,
+                0.781
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "This paper introduces BANGLABOOK, the largest Bangla book review dataset with 158,065 samples, each labeled with 1 of 3 user sentiments. We provide extensive statistical analysis and strong baselines facilitating the utility of the dataset. Given its massive size and fine-grained sentiment distribution, BANGLABOOK has the potential to alleviate the resource scarcity in Bangla language research."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1241"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.251,
+                0.099
+            ],
+            "angle": 0,
+            "content": "7 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.11,
+                0.493,
+                0.368
+            ],
+            "angle": 0,
+            "content": "Many of the reviews that were gathered for constructing BANGLABOOK are discarded because they lack a corresponding rating. A manual annotation process would have yielded a much larger dataset, which was not feasible due to resource constraints. Moreover, one of the challenges for validating the dataset is the lack of statistical models and word-embeddings pre-trained on the Bangla language. Some pre-trained Bangla-BERT models, yet to be trained on extensive corpora, have only recently been proposed. Improving transformer-based models for Bangla can enhance sub-word level contextual understanding which will consequently help in more accurate identification of the sentiments in BANGLABOOK (Islam et al., 2022)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.085,
+                0.61,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.108,
+                0.885,
+                0.189
+            ],
+            "angle": 0,
+            "content": "Mst Tuhin Akter, Manoara Begum, and Rashed Mustafa. 2021. Bengali sentiment analysis of e-commerce product reviews using k-nearest neighbors. In 2021 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD), pages 40-44. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.198,
+                0.887,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Sanjida Akter and Muhammad Tareq Aziz. 2016. Sentiment analysis on facebook group using lexicon based approach. In 2016 3rd International Conference on Electrical Engineering and Information Communication Technology (ICEEICT), pages 1-4. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.288,
+                0.887,
+                0.368
+            ],
+            "angle": 0,
+            "content": "Shad Al Kaiser, Sudipta Mandal, Ashraful Kalam Abid, Ekhfa Hossain, Ferdous Bin Ali, and Intisar Tahmid Naheen. 2021. Social media opinion mining based on bangla public post of facebook. In 2021 24th International Conference on Computer and Information Technology (ICCIT), pages 1-6. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.378,
+                0.886,
+                0.497
+            ],
+            "angle": 0,
+            "content": "Abhik Bhattacharjee, Tahmid Hasan, Wasi Ahmad, Kazi Samin Mubasshir, Md Saiful Islam, Anindya Iqbal, M. Sohel Rahman, and Rifat Shahriyar. 2022. BanglaBERT: Language model pretraining and benchmarks for low-resource language understanding evaluation in Bangla. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 1318-1327, Seattle, United States. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.508,
+                0.885,
+                0.534
+            ],
+            "angle": 0,
+            "content": "L. Breiman. 2001. Random forests. Machine Learning, 45:5-32."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.545,
+                0.885,
+                0.6
+            ],
+            "angle": 0,
+            "content": "Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.609,
+                0.885,
+                0.675
+            ],
+            "angle": 0,
+            "content": "Shaika Chowdhury and Wasifa Chowdhury. 2014. Performing sentiment analysis in bangla microblog posts. In 2014 International Conference on Informatics, Electronics & Vision (ICIEV), pages 1-6. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.686,
+                0.885,
+                0.726
+            ],
+            "angle": 0,
+            "content": "Corinna Cortes and Vladimir Naumovich Vapnik. 1995. Support-vector networks. Machine Learning, 20:273-297."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.737,
+                0.885,
+                0.791
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.801,
+                0.885,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Rajib Chandra Dey and Orvila Sarker. 2019. Sentiment analysis on bengali text using lexicon based approach. In 2019 22nd International Conference on Computer and Information Technology (ICCIT), pages 1-5. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.878,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Asif Hassan, Mohammad Rashedul Amin, Abul Kalam Al Azad, and Nabeel Mohammed. 2016. Sentiment analysis on bangla and romanized bangla text using"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.108,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1242"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.489,
+                0.125
+            ],
+            "angle": 0,
+            "content": "deep recurrent models. In 2016 International Workshop on Computational Intelligence (IWCI), pages 51-56. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.138,
+                0.489,
+                0.177
+            ],
+            "angle": 0,
+            "content": "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9:1735-1780."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.19,
+                0.489,
+                0.243
+            ],
+            "angle": 0,
+            "content": "MD Iqbal, Avishek Das, Omar Sharif, Mohammed Moshiul Hoque, and Iqbal H Sarker. 2022. Bemoc: A corpus for identifying emotion in bengali texts. SN Computer Science, 3(2):1-17."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.255,
+                0.489,
+                0.333
+            ],
+            "angle": 0,
+            "content": "Khondoker Ittehadul Islam, Md Saiful Islam, and Md Ruhul Amin. 2020. Sentiment analysis in bengali via transfer learning using multi-lingual bert. In 2020 23rd International Conference on Computer and Information Technology (ICCIIT), pages 1-5. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.346,
+                0.489,
+                0.412
+            ],
+            "angle": 0,
+            "content": "Khondoker Ittehadul Islam, Sadipta Kar, Md Saiful Islam, and Mohammad Ruhul Amin. 2021. Sentnob: A dataset for analysing sentiment on noisy bangla texts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3265-3271."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.424,
+                0.489,
+                0.53
+            ],
+            "angle": 0,
+            "content": "Khondoker Ittehadul Islam, Tanvir Yuvraz, Md Saiful Islam, and Enamul Hassan. 2022. Emonoba: A dataset for analyzing fine-grained emotions on noisy bangla texts. 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, pages 128-134."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.541,
+                0.489,
+                0.581
+            ],
+            "angle": 0,
+            "content": "George H. John and Pat Langley. 1995. Estimating continuous distributions in bayesian classifiers. In _Conference on Uncertainty in Artificial Intelligence_."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.593,
+                0.489,
+                0.659
+            ],
+            "angle": 0,
+            "content": "Mst Eshita Khatun and Tapasy Rabeya. 2022. A machine learning approach for sentiment analysis of book reviews in bangla language. In 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), pages 1178-1182. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.671,
+                0.489,
+                0.71
+            ],
+            "angle": 0,
+            "content": "Saskia le Cessie and J. C. van Houwelingen. 1992. Ridge estimators in logistic regression. Applied statistics, 41:191-201."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.722,
+                0.489,
+                0.776
+            ],
+            "angle": 0,
+            "content": "Tom MY Lin, Pin Luarn, and Yun Kuei Huang. 2005. Effect of internet book reviews on purchase intention: A focus group study. The Journal of Academic Librarianship, 31(5):461-468."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.787,
+                0.489,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Shamsul Arafin Mahtab, Nazmul Islam, and Md Mahfuzur Rahaman. 2018. Sentiment analysis on bangladesh cricket with support vector machine. In 2018 international conference on Bangla speech and language processing (ICBSLP), pages 1-4. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "P Majumder, M Mitra, and BB Chaudhuri. 2002. Ngram: a language independent approach to ir and nlp. In International conference on universal knowledge and language, volume 2."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Muhammad Mahmudun Nabi, Md Tanzir Altaf, and Sabir Ismail. 2016. Detecting sentiment from bangla text using machine learning technique and feature analysis. International Journal of Computer Applications, 153(11):28-34."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.16,
+                0.883,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.221,
+                0.883,
+                0.286
+            ],
+            "angle": 0,
+            "content": "Md Rahib, Rumman Hussain Khan, Amzad Hussain Tamim, Mohammad Zawad Tahmeed, and Mohammad Jaber Hossain. 2022. Emotion detection based on bangladeshi peoples social media response on Covid-19. SN Computer Science, 3(2):1-6."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.295,
+                0.883,
+                0.373
+            ],
+            "angle": 0,
+            "content": "Fuad Rahman, Habibur Khan, Zakir Hossain, Mahfuza Begum, Sadia Mahanaz, Ashraful Islam, and Aminul Islam. 2019. An annotated bangla sentiment analysis corpus. In 2019 International Conference on Bangla Speech and Language Processing (ICBSLP), pages 1-5. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.381,
+                0.883,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Md Atikur Rahman and Emon Kumar Dey. 2018. Datasets for aspect-based sentiment analysis in bangla and its baseline evaluation. Data, 3(2):15."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.429,
+                0.883,
+                0.469
+            ],
+            "angle": 0,
+            "content": "Gillian Catriona Ramchand. 2004. Two types of negation in bengali. *Clause structure in South Asian languages*, pages 39-66."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.476,
+                0.883,
+                0.504
+            ],
+            "angle": 0,
+            "content": "Sagor Sarker. 2020. Banglabert: Bengali mask language model for bengali language understanding."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.511,
+                0.883,
+                0.564
+            ],
+            "angle": 0,
+            "content": "Salim Sazzed. 2020a. Cross-lingual sentiment classification in low-resource bengali language. In Proceedings of the sixth workshop on noisy user-generated text (W-NUT 2020), pages 50–60."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.572,
+                0.883,
+                0.637
+            ],
+            "angle": 0,
+            "content": "Salim Sazzed. 2020b. Development of sentiment lexicon in bengali utilizing corpus and cross-lingual resources. In 2020 IEEE 21st International conference on information reuse and integration for data science (IRI), pages 237-244. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.645,
+                0.883,
+                0.698
+            ],
+            "angle": 0,
+            "content": "Salim Sazzed. 2021. Bengsentilex and bengswear-lex: creating lexicons for sentiment analysis and profanity detection in low-resource bengali language. PeerJ Computer Science, 7:e681."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.706,
+                0.883,
+                0.771
+            ],
+            "angle": 0,
+            "content": "Marina Sokolova, Nathalie Japkowicz, and S. Szpakowicz. 2006. Beyond accuracy, f-score and roc: A family of discriminant measures for performance evaluation. In Australian Conference on Artificial Intelligence."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.779,
+                0.883,
+                0.832
+            ],
+            "angle": 0,
+            "content": "Alan T Sorensen and Scott J Rasmussen. 2004. Is any publicity good publicity? a note on the impact of book reviews. NBER Working paper, Stanford University."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.84,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Nusrath Tabassum and Muhammad Ibrahim Khan. 2019. Design an empirical framework for sentiment analysis from bangla text using machine learning. In 2019 International Conference on Electrical, Computer and Communication Engineering (ECCE), pages 1-5. IEEE."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1243"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.152
+            ],
+            "angle": 0,
+            "content": "SM Abu Taher, Kazi Afsana Akhter, and KM Azharul Hasan. 2018. N-gram based sentiment mining for bangla text using support vector machine. In 2018 international conference on Bangla speech and language processing (ICBSLP), pages 1-5. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.162,
+                0.487,
+                0.201
+            ],
+            "angle": 0,
+            "content": "Hanne-Ruth Thompson. 2006. Negation patterns in bengali. Bulletin of the School of Oriental and African Studies, 69(2):243-265."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.212,
+                0.487,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Rashedul Amin Tuhin, Bechitra Kumar Paul, Faria Nawrine, Mahbuba Akter, and Amit Kumar Das. 2019. An automated system of sentiment analysis from bangla text using supervised learning techniques. In 2019 IEEE 4th International Conference on Computer and Communication Systems (ICCCS), pages 360-364. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.313,
+                0.487,
+                0.365
+            ],
+            "angle": 0,
+            "content": "Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. ArXiv, abs/1706.03762."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.376,
+                0.487,
+                0.454
+            ],
+            "angle": 0,
+            "content": "Anning Wang, Qiang Zhang, Shuangyao Zhao, Xiaonong Lu, and Zhanglin Peng. 2020. A review-driven customer preference measurement model for product improvement: sentiment-based importance performance analysis. Information Systems and e-Business Management, 18:61-88."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.465,
+                0.487,
+                0.505
+            ],
+            "angle": 0,
+            "content": "Alice Zheng and Amanda Casari. 2018. Feature engineering for machine learning: principles and techniques for data scientists. \"O'Reilly Media, Inc.\""
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.505
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.518,
+                0.237,
+                0.534
+            ],
+            "angle": 0,
+            "content": "A Appendix"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1244"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.115,
+                0.143,
+                0.932,
+                0.759
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td colspan=\"2\">Sentiment Classification</td><td colspan=\"2\">Sentiment Distribution</td><td>Total # of Samples</td><td>Availability</td><td>Type of Content</td><td>Source(s)</td><td>Baseline Models</td></tr><tr><td>(Tabassum and Khan, 2019)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>1,050</td><td>Closed</td><td>Posts, comments</td><td>Facebook, Twitter</td><td>RF</td></tr><tr><td>(Chowdhury and Chowdhury, 2014)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>1,300</td><td>Closed</td><td>Posts, comments</td><td>Twitter</td><td>-</td></tr><tr><td>(Nabi et al., 2016)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>1,500</td><td>Closed</td><td>Posts, comments</td><td>Social Media</td><td>-</td></tr><tr><td>(Mahtab et al., 2018)</td><td colspan=\"2\">Praise Criticism Sadness</td><td colspan=\"2\">513 604 484</td><td>1,601</td><td>Closed</td><td>Comments</td><td>Prothom Alo Online News Portal</td><td>SVM, DT, NB</td></tr><tr><td>(Akter and Aziz, 2016)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">- -</td><td>3,600</td><td>Closed</td><td>Posts, comments</td><td>Facebook</td><td>NB</td></tr><tr><td>(Rahman and Kumar Dey, 2018)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">- -</td><td>4,700</td><td>Open</td><td>Comments</td><td>Facebook pages: BBC Bangla, Prothom Alo</td><td>SVM, LR, KNN, DT, LSTM NB, CNN</td></tr><tr><td>(Dey and Sarker, 2019)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">2,600 2,600</td><td>5,200</td><td>Closed</td><td>Comments, reviews</td><td>Facebook, Twitter, YouTube, News Portals</td><td>DT, NB, SVM</td></tr><tr><td>(Khatun and Rabeya, 2022)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>5,500</td><td>Closed</td><td>Comments, reviews</td><td>Social Media</td><td>-</td></tr><tr><td>BEMOC (Iqbal et al., 2022)</td><td colspan=\"2\">AngerFearSadJoyDisgust</td><td colspan=\"2\">- -</td><td>7,000</td><td>Open</td><td>Posts, comments</td><td>Facebook, YouTube, Online blogs, Bangla story books, novels, newspapers, discourse</td><td>-</td></tr><tr><td>(Tuhin et al., 2019)</td><td colspan=\"2\">Happy Tender Excited Sad Angry Scared</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>(Akter et al., 2021)</td><td colspan=\"2\">Positive Negative Neutral</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td>7,905</td><td>Closed</td><td>Product reviews</td><td>Daraz</td><td>RF, LR, SVM, KNN, XGB</td><td></td></tr><tr><td>(Hassan et al., 2016)</td><td colspan=\"2\">Positive Negative Ambiguous</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>(Rahib et al., 2022)</td><td colspan=\"2\">Insightful Curious Gratitude</td><td colspan=\"2\">3,800 3,549 3,232</td><td>10,581</td><td>Closed</td><td>Comments</td><td>Social Media</td><td>SVM, RF, CNN, LSTM</td></tr><tr><td rowspan=\"2\">(Al Kaiser et al., 2021)</td><td rowspan=\"2\">Positive Negative Neutral</td><td>Wishful Thinking Appreciation</td><td>967 942</td><td>1,909</td><td rowspan=\"2\">11,006</td><td rowspan=\"2\">Closed</td><td rowspan=\"2\">Comments</td><td rowspan=\"2\">Facebook</td><td rowspan=\"2\">LR, DT, RF, MNB, KNN, Linear SVM, RBF SVM, XGB</td></tr><tr><td>Gender-based hate Religious hate Political hate Personal hate Sarcasm</td><td>525 731 572 1,995 1,414 3,860</td><td>5,237 3,860</td></tr><tr><td>(Sazzed, 2020a)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">8,500 3,307</td><td>11,807</td><td>Open</td><td>Comments</td><td>YouTube</td><td>SVM, ET, RF, LR, VADER, TextBlob</td></tr><tr><td>(Sazzed, 2020b)</td><td colspan=\"2\">Positive Negative</td><td colspan=\"2\">- -</td><td>12,000</td><td>Closed</td><td>Comments</td><td>YouTube</td><td>-</td></tr><tr><td rowspan=\"2\">SENTNoB (Islam et al., 2021)</td><td rowspan=\"2\" colspan=\"2\">Positive Negative</td><td rowspan=\"2\" colspan=\"2\">6,410 5,709 3,609</td><td rowspan=\"2\">15,728</td><td rowspan=\"2\">Open</td><td rowspan=\"2\">Comments</td><td rowspan=\"2\">Prothom Alo Online Newspaper, YouTube</td><td rowspan=\"2\">RNN</td></tr><tr></tr><tr><td>(Islam et al., 2020)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">4,769 8,351 4,732</td><td>17,852</td><td>Open</td><td>Comments</td><td>Prothom Alo Online Newspaper</td><td>CNN, LSTM, BERT, GRU, fastText</td></tr><tr><td>EMONOBA (Islam et al., 2022)</td><td colspan=\"2\">Love Joy Surprise Anger Sadness Fear</td><td colspan=\"2\">4,202 9,249 939 3,905 5,109 307</td><td>20,468</td><td>Open</td><td>Comments</td><td>YouTube, Facebook, Twitter, Prothom Alo</td><td>Bi-LSTM, fastText, Bangla-BERT-base</td></tr><tr><td>BANGLABOOK (ours)</td><td colspan=\"2\">Positive Negative Neutral</td><td colspan=\"2\">141,587 9,674 6,804</td><td>158,065</td><td>\\( Open^† \\)</td><td>Book reviews</td><td>Rokomari, Wafilife</td><td>RF, LSTM, LR, GRU MNB, SVM, XGB Bangla-BERT</td></tr></table>"
+        },
+        {
+            "type": "table_footnote",
+            "bbox": [
+                0.113,
+                0.761,
+                0.885,
+                0.863
+            ],
+            "angle": 0,
+            "content": "Table 5: Comparison of notable Bangla Sentiment Analysis datasets sorted in ascending order of size. The abbreviations respectively denote, Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), Naive Bayes (NB), Logistic Regression (LR), K-Nearest Neighbors (KNN), Long Short-term Memory (LSTM), Convolutional Neural Network (CNN), Extreme Gradient Boost (XGB), Multinomial Naive Bayes (MNB), Radial Basis Function (RBF), Extreme Random Tree (ET), Recurrent Neural Network (RNN), Bidirectional Encoder Representations from Transformers (BERT), Gated Recurrent Unit (GRU), Bi-LSTM (Bidirectional LSTM). All the publicly available datasets are hyperlinked. Open† denotes the redaction of the link for anonymity."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1245"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.133,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.127,
+                0.533,
+                0.144
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.145,
+                0.233,
+                0.16
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.171,
+                0.553,
+                0.187
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.188,
+                0.351,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.213,
+                0.696,
+                0.23
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.231,
+                0.233,
+                0.246
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.257,
+                0.669,
+                0.273
+            ],
+            "angle": 0,
+            "content": "A4. Have you used AI writing assistants when working on this paper?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.274,
+                0.233,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.301,
+                0.49,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.322,
+                0.215,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.348,
+                0.53,
+                0.364
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.365,
+                0.351,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.391,
+                0.779,
+                0.407
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.408,
+                0.351,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.434,
+                0.882,
+                0.497
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.498,
+                0.351,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.525,
+                0.882,
+                0.573
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.574,
+                0.351,
+                0.589
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.6,
+                0.881,
+                0.632
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.633,
+                0.351,
+                0.648
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.658,
+                0.882,
+                0.74
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.741,
+                0.233,
+                0.755
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.765,
+                0.494,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.787,
+                0.215,
+                0.802
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.814,
+                0.882,
+                0.846
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.847,
+                0.233,
+                0.861
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.869,
+                0.878,
+                0.893
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1246"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.085,
+                0.88,
+                0.133
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.144,
+                0.881,
+                0.207
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.218,
+                0.881,
+                0.283
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.13,
+                0.085,
+                0.881,
+                0.283
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.294,
+                0.878,
+                0.331
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.341,
+                0.881,
+                0.388
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.4,
+                0.881,
+                0.464
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.476,
+                0.881,
+                0.54
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.55,
+                0.873,
+                0.582
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.593,
+                0.881,
+                0.642
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.128,
+                0.341,
+                0.881,
+                0.642
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1247"
+        }
+    ]
+]

2023/BanglaBook_ A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews/28378eaf-0a48-458e-a698-291e5038e7bf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59c4837a19f13dfb34918e4f519caeba735bc74c7870acad9b087e833130451c
+size 502758

2023/BanglaBook_ A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews/full.md

@@ -0,0 +1,259 @@
+# BANGLABOOK: A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews
+
+Mohsinul Kabir*, Obayed Bin Mahfuz*, Syed Rifat Raiyan*, Hasan Mahmud, Md Kamrul Hasan
+
+Systems and Software Lab (SSL)
+
+Department of Computer Science and Engineering
+
+Islamic University of Technology, Dhaka, Bangladesh
+
+{mohsinulkabir, siam, rifatraiyan, hasan, hasank}@iut-dhaka.edu
+
+# Abstract
+
+The analysis of consumer sentiment, as expressed through reviews, can provide a wealth of insight regarding the quality of a product. While the study of sentiment analysis has been widely explored in many popular languages, relatively less attention has been given to the Bangla language, mostly due to a lack of relevant data and cross-domain adaptability. To address this limitation, we present BANGLABOOK, a large-scale dataset of Bangla book reviews consisting of 158,065 samples classified into three broad categories: positive, negative, and neutral. We provide a detailed statistical analysis of the dataset and employ a range of machine learning models to establish baselines including SVM, LSTM, and Bangla-BERT. Our findings demonstrate a substantial performance advantage of pretrained models over models that rely on manually crafted features, emphasizing the necessity for additional training resources in this domain. Additionally, we conduct an in-depth error analysis by examining sentiment unigrams, which may provide insight into common classification errors in under-resourced languages like Bangla. Our codes and data are publicly available at https://github.com/mohsinulkabir14/BanglaBook.
+
+# 1 Introduction
+
+The resources publicly available for scholarly investigation in the realm of Sentiment Analysis (SA) for the Bangla language are scarce and limited in quantity (Khatun and Rabeya, 2022; Sazzed, 2021; Rahman et al., 2019) despite its literary gravitas as the  $6^{\text{th}}$  most spoken language $^{1}$  in the world with approximately 200 million speakers. In the existing literature on Bangla Text SA, as shown in Table 5, the largest dataset consists
+
+of 20,468 samples (Islam et al., 2022) while the smallest has a mere 1,050 samples (Tabassum and Khan, 2019). Besides these, Islam et al. (2020) created a dataset consisting of 17,852 samples and Islam et al. (2021) utilized a dataset of 15,728 samples. All other datasets apart from these either have  $< 15,000$  samples or are publicly unavailable. Another limitation of the existing research works in Bangla Text SA is the deficiency of datasets having product-specific review samples. Most of the available Bangla SA datasets are focused on user-generated textual content from cyberspace. The insights derived from these may not accurately represent sentiment in the context of product reviews, thus hindering their usefulness for businesses. The tonal and linguistic analysis of reviews from product-specific datasets can aid businesses to gain valuable insights into customer attitudes, preferences, and experiences which can then be leveraged to improve products and services, design targeted marketing campaigns, and make more informed business decisions. In this paper, we introduce a large-scale dataset, BANGLABOOK, consisting of 158,065 samples of book reviews collected from online bookshops written in the Bangla language. This is the largest dataset for Bangla sentiment analysis to the best of our knowledge. We perform an analysis of the dataset's statistical characteristics, employ various ML techniques to establish a performance benchmark for validating the dataset, and also conduct a thorough evaluation of the classification errors.
+
+# 2 Dataset Construction
+
+In order to create this dataset, we collect a total of 204,659 book reviews from two online bookshops (Rokomari² and Wafilife³) using a web scraper developed with several Python libraries, including BeautifulSoup, Selenium, Pandas, Openpyxl,
+
+$^{2}$ https://www.rokomari.com/  
+<sup>3</sup>https://www.wafilife.com/
+
+<table><tr><td>Source</td><td>Language</td><td>Annotated</td><td>Unannotated</td><td>Total</td></tr><tr><td rowspan="4">Rokomari</td><td>Bangla</td><td>84,672</td><td>24,806</td><td>109,478</td></tr><tr><td>Bangla† + English</td><td>56,485</td><td>15,408</td><td>71,893</td></tr><tr><td>Bangla + Bangla† + English</td><td>13,144</td><td>4,114</td><td>17,258</td></tr><tr><td>Total</td><td>154,301</td><td>44,328</td><td>198,629</td></tr><tr><td rowspan="4">Wafilife</td><td>Bangla</td><td>4,699</td><td>59</td><td>4,758</td></tr><tr><td>Bangla† + English</td><td>370</td><td>3</td><td>373</td></tr><tr><td>Bangla + Bangla† + English</td><td>896</td><td>3</td><td>899</td></tr><tr><td>Total</td><td>5,965</td><td>65</td><td>6,030</td></tr><tr><td rowspan="4"></td><td>Bangla</td><td>89,371</td><td>24,865</td><td>114,237</td></tr><tr><td>Bangla† + English</td><td>56,855</td><td>15,411</td><td>72,266</td></tr><tr><td>Bangla + Bangla† + English</td><td>14,040</td><td>4,117</td><td>18,157</td></tr><tr><td>Total</td><td>160,266</td><td>44,393</td><td>204,659</td></tr><tr><td></td><td>Untranslated Data (Removed)</td><td>2,201</td><td></td><td></td></tr><tr><td></td><td>Final Dataset Size</td><td>158,065</td><td></td><td></td></tr></table>
+
+and Webdriver, to collect and process the raw data.
+
+For the data collection and preparation process of the BANGLABOOK dataset, we first compile a list of URLs for authors from online bookstores. From there, we procure URLs for the books. We meticulously scrape information such as book titles, author names, book categories, review texts, reviewer names, review dates, and ratings by utilizing these book URLs.
+
+Table 1: Summary statistics of our dataset. Bangla† denotes Romanized Bangla text.  
+
+<table><tr><td rowspan="5">General Properties</td><td># of Reviews</td><td>158,065</td></tr><tr><td># of Books</td><td>30,253</td></tr><tr><td># of Reviewers</td><td>44,616</td></tr><tr><td># of Categories</td><td>1,573</td></tr><tr><td>Total Review Words</td><td>44,429,201</td></tr><tr><td rowspan="3">Single Review</td><td>max. Word length</td><td>765</td></tr><tr><td>min. Word length</td><td>1</td></tr><tr><td>avg. Word length</td><td>281.08</td></tr></table>
+
+Table 2: General overview of BANGLABOOK.
+
+# 2.1 Labeling & Translation
+
+If a review does not have a rating, we deem it unannotated. Reviews with a rating of 1 or 2 are classified as negative, a rating of 3 is considered neutral, and a rating of 4 or 5 is classified as positive. Two manual experiments are carried out to validate the use of ratings as a measure of sentiment in product reviews. In the first experiment, around  $10\%$  of the reviews are randomly selected and annotated manually. The annotated labels are cross-checked with the original labels, resulting in a  $96.7\%$  accuracy in the corresponding labels. In addition, we consult the work of Wang et al. (2020) that explored the issue of incongruous sentiment expressions with regard to ratings. Specifically, the study scrutinized two categories of reviews: high ratings lacking a positive sentiment, and low ratings lacking a negative sentiment. We perform an analysis to identify such inconsistencies within our dataset and discovered that only a minuscule  $3.41\%$  of the samples exhibited this pattern. This figure is rela
+
+tively insignificant when considering the substantially large scale of our dataset.
+
+![](images/76550d660072eace8cce26377b4dcae393302544a68c254e887f398292922d2d.jpg)  
+(a) Sentiment Distribution
+
+![](images/21e6fd1cf1fee406816d2a9d897267b82959eb266caa2730ad1e1353cc05df59.jpg)  
+(b) Rating Distribution  
+Figure 1: Class Distribution of BANGLABOOK.
+
+After discarding the unannotated reviews, we curate a final dataset of 158,065 annotated reviews. Of these, 89,371 are written entirely in Bangla. The remaining 68,694 reviews were written in Romanized Bangla, English, or a mix of languages. They are translated into Bangla with Google Translator and a custom Python program using the googletrans library. The translations are subsequently subjected to manual review and scrutiny to confirm their accuracy. The majority of inaccurate translations primarily comprise spelling errors and instances where English words remain untranslated within samples containing a combination of Bangla and English text. The meticulous evaluation process of untranslated samples involves a thorough assessment by post-graduate native Bangla speakers, who critically compare the translated text against the original untranslated text to ascertain the correctness of the translation.
+
+# 3 Statistical Analysis
+
+Tables 1 and 2 provide an overview of the statistical properties of the BANGLABOOK dataset. The sentiment doughnut chart in Figure-1a illustrates the proportion of positive, neutral, and negative reviews, while the rating doughnut chart in Figure-1b displays the percentage of reviews that corre
+
+<table><tr><td>Method</td><td>Negative</td><td>Neutral</td><td>Positive</td><td>Weighted Avg.</td></tr><tr><td>Random Forest (word 2-gram + word 3-gram)</td><td>0.56</td><td>0.34</td><td>0.96</td><td>0.9106</td></tr><tr><td>SVM (word 2-gram + word 3-gram)</td><td>0.40</td><td>0.15</td><td>1.00</td><td>0.9053</td></tr><tr><td>Random Forest (word 1-gram)</td><td>0.48</td><td>0.35</td><td>0.96</td><td>0.9043</td></tr><tr><td>Logistic Regression (char 2-gram + char 3-gram)</td><td>0.55</td><td>0.13</td><td>0.96</td><td>0.8978</td></tr><tr><td>Bangla-BERT (base-uncased)</td><td>0.60</td><td>0.22</td><td>0.96</td><td>0.9064</td></tr><tr><td>Logistic Regression (word 2-gram + word 3-gram)</td><td>0.53</td><td>0.13</td><td>0.96</td><td>0.8964</td></tr><tr><td>Bangla-BERT (large)</td><td>0.72</td><td>0.40</td><td>0.97</td><td>0.9331</td></tr><tr><td>XGBoost (char 2-gram + char 3-gram)</td><td>0.31</td><td>0.02</td><td>0.95</td><td>0.8723</td></tr><tr><td>Multinomial NB (word 2-gram + word 3-gram)</td><td>0.23</td><td>0.03</td><td>0.95</td><td>0.8663</td></tr><tr><td>LSTM (GloVe)</td><td>0.11</td><td>0.00</td><td>0.10</td><td>0.0991</td></tr><tr><td>XGBoost (word 2-gram + word 3-gram)</td><td>0.23</td><td>0.01</td><td>0.95</td><td>0.8651</td></tr><tr><td>Multinomial NB (BoW)</td><td>0.18</td><td>0.05</td><td>0.94</td><td>0.8564</td></tr><tr><td>SVM (word 1-gram)</td><td>0.08</td><td>0.04</td><td>0.94</td><td>0.8519</td></tr></table>
+
+Table 3: Catergory-wise Binary Task F1-score and Weighted Average F1-score of each method on BANGLABOOK.
+
+spond to each rating on a scale of 1 to 5.
+
+Upon analyzing the sentiment chart, it appears that the majority of the reviews  $(124,084 + 17,503 = 141,587$  samples) are positive, with a significant portion also being negative  $(2,728 + 6,946 = 9,674$  samples). A relatively small fraction of the reviews are neutral (6,804 samples). This suggests that overall, the books have been well received by the readers, with the majority expressing favorable opinions. The distribution of the dataset is representative of real-world scenarios and it tessellates well with previous content analysis works on book reviews (Lin et al., 2005; Sorensen and Rasmussen, 2004). In Figure-2, we can visualize an illustration of the sentiment distribution among the 5 most frequently reviewed categories of books. We can gain some salient insights from the popularity of these genres. Contemporary novels are bestsellers as they reflect current events, social issues, and trends, making them relatable and thought-provoking for the readers while self-help and religious books provide guidance, inspiration, and a sense of purpose, catering to individuals' quest for personal growth and spiritual fulfillment.
+
+# 4 Developing Benchmark for BANGLABOOK
+
+A series of baseline models with combinations of different lexical and semantic features are chosen to evaluate the BANGLABOOK dataset. An overview of the models, evaluation metrics, results, and analysis of the experimental results are provided in this section.
+
+# 4.1 Baseline Models & Features
+
+For the lexical features, we extract bag-of-words (BoW), char  $n$ -grams (1-3), and word  $n$ -grams (1-3) from the reviews as these representations have performed well in different classification tasks (Islam et al., 2022). After extracting the features, they are vectorized using TF-IDF and count vec
+
+![](images/801e495ea340c69d691c05085acf634158a1c49e822d0ae50a93be87c223f433.jpg)  
+Figure 2: Sentiment Distribution of top 5 most popular genres. In clockwise order, वस्थकालीन अप्रति स (Contemporary Novel), वस्थकालीन अप्रति तूधोत्तिक तूधोत्तिक तूधोत्तिक तूधोत्तिक तूधोत्तिक (Islamic Ideals and Doctrines), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक तूधोत्तिक (Islamic Books: Self-Development), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक (Student Life Development), वस्थकालीन वर्षेहि तूधोत्तिक तूधोत्तिक (Translated Books: Self-Development and Meditation).
+
+torizer and trained on a series of ML models such as Random Forest (Breiman, 2001), XGBoost (Chen and Guestrin, 2016), linear SVM (Cortes and Vapnik, 1995), Logistic Regression (le Cessie and van Houwelingen, 1992) and Multinomial Naive Bayes (John and Langley, 1995). We choose LSTM (Hochreiter and Schmidhuber, 1997) with GloVe (Pennington et al., 2014) embedding for its ability to understand context along with recent dependency. We also fine-tuned two available transformer-based models in Bangla: Bangla-BERT (base-uncased) (110M parameters) (Sarker, 2020) and Bangla-BERT (large) (2.5B parameters) (Bhattacharjee et al., 2022), due to the recent success of BERT (Devlin et al., 2019) in various downstream NLP tasks. We select F1-score and weighted average F1-score to evaluate the models because the dataset has an un
+
+<table><tr><td>Class</td><td>10 Most Frequent Words</td></tr><tr><td>Positive</td><td>बालेला (good), शन्तर (nice), अस्थान्यारुँ (extraordinary), दूपाकोँ (splendid), वर्णा (best), सहानुँ (facile), धनाकोर (beautiful), अश्थि (hope), वन्तर्यादित (gratefulness), अलामान्यारुँ, वल्तर्यादित (gratitude)</td></tr><tr><td>Neutral</td><td>बालेला (good), शन्तर (nice), अशरागोँ (bad), वन्तार्यादित (average), वल (fault), वर्णारेम (hope), सहानुँ (facile), धनाकोर (extraordinary), कम्ज (low), वन्तर्यादित (splendid)</td></tr><tr><td>Negative</td><td>बालेला (good), वल्तर (trash), अशरागोँ (bad), वल (fault), वर्णारेम (nice), वर्णारेम (hope), वति (waste), कम्ज (useless), वन्तर्यादित (disappointment), कम्ज (low)</td></tr></table>
+
+Table 4: Most frequent word unigrams conveying the strongest sentiments of each class with English translation. The colors respectively denote Positive, Neutral and Negative sentiments.
+
+![](images/e56d3518d85c08bfa6d3d061bdfe7945c590061f3d16e56b768f41c191a92c4f.jpg)  
+Figure 3: Confusion Matrix for Bangla-BERT
+
+even class distribution. F1-score is the harmonic mean of precision and recall and it helps balance the metric across the imbalanced positive/negative samples (Sokolova et al., 2006). All our experiments are done using scikit-learn, pytorch, and transformers (Vaswani et al., 2017) and run on Google Colaboratory. The training, testing, and validation split of the entire dataset was 70-20-10 with previously unseen samples in the test and validation set.
+
+# 4.2 Results & Findings
+
+Table 3 summarizes the experimental results for BANGLABOOK. Results show that BanglaBERT (large) outperforms all other models by a clear margin. Also, the combination of word/char 2-gram and word/char 3-gram perform exceptionally well with respective classifiers. Our hypothesis is that these two features result in a large number of unique word and character combinations, aiding the models' ability to generalize effectively across categories. Islam et al. (2022, 2021); Taher et al. (2018) concur on the same verdict by implying that the task predominantly relies on word units, with minimal dependence on subword level information and the nature of the Bangla language itself. Furthermore, Majumder et al. (2002) outlined the suitability of  $n$ -gram approaches in generating language profiles for Indo
+
+European languages in their work. The bag-of-words (BoW) feature is inept at classifying the corresponding categories because of its inability to capture critical contextual information and nuance (Zheng and Casari, 2018). Although word 1-gram does not outperform word 2-gram and word 3-gram, it does predict the 'Neutral' class well. Both the pre-trained Bangla-BERT models perform fairly consistently across all categories on the BANGLABOOK dataset, demonstrating the usefulness of contextual understanding and transfer learning in classification tasks even in low-resource languages like Bangla. The LSTM model with GloVe embedding recognizes the 'Negative' and 'Positive' classes very marginally and fails completely to identify the 'Neutral' category. It is also notable that, SVM with bigram and trigram achieves perfect scores in the 'Positive' class.
+
+To summarize, the utilization of pre-trained models (i.e. Bangla-BERT) that undergo training on extensive corpora, leading to exposure to extensive general language knowledge, has significantly contributed to their superior classification performance compared to other models and word embeddings. Additionally, models trained on handcrafted features also perform significantly well. It should be noted that Bangla pre-trained models are currently undergoing development, and further training on expansive corpora has the potential to enhance their ability to generalize and achieve even more impressive results.
+
+# 4.3 Error Analysis
+
+In the 'Positive' class, all the models produce excellent classification results. While some models perform reasonably well on the 'Negative' class, nearly all of the models perform poorly on the 'Neutral' class. The class imbalance of the dataset, as shown in Figure 1, is one obvious cause of this fluctuation in results. The confusion matrix for
+
+Bangla-BERT on our dataset, presented in Figure3, reveals that most of the 'Negative' and 'Neutral' samples are misclassified as 'Positive' samples by our classifiers. To further analyze the misclassifications, we examine the W1 (word unigrams) of these three classes. We find 124,796 unique W1 for the 'Positive' class, 20,714 unique W1 for the 'Negative' class, and 19,096 unique W1 for the 'Neutral' class.  $77.57\%$  of the W1 from the 'Neutral' class and  $79.83\%$  of the W1 from the 'Negative' class are found in the 'Positive' class. Table 4 depicts the most frequent W1 conveying the strongest sentiments for each class. With only one distinct 'Neutral' W1 and even the 'Negative' class having multiple positive W1, the dominance of 'Positive' sentiment W1 over the other two classes is evident. This may have contributed to the lack of distinctive words in the 'Negative' and 'Neutral' classes, which inevitably prevented the feature-based models from generalizing.
+
+# 5 Morphology and Negation Patterns of Bangla
+
+Understanding the morphology and negation patterns of a language holds paramount importance in the realm of sentiment analysis because negation can alter the meaning of words and phrases, thereby affecting the overall sentiment conveyed by a text. We provide a concise yet insightful recapitulation of the topic in the case of Bangla accompanied by review samples from our dataset BANGLABOOK as the respective examples. From the linguistic typological standpoint, Bangla is categorized as a subject-object-verb (SOV) language because the subject, object, and verb generally adhere to said order in its sentential structure (Ramchand, 2004). The most common juxtaposition of polarity from positive to negative is the use of  $n_i$  ( $\widehat{n}$ ) as a tensed negative. For example,
+
+Aaannn aonnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnae ananrnnnnaee
+
+Another negational feature is expressed by placing na (nata) prior to the non-finite verb and after the finite verb in a sentence (although there are some exceptions). For example,
+
+AeBnBnBn BnBnBnBn BnBnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn BnBn
+
+The Bangla language consists of no negative adverbs or pronouns (Thompson, 2006). This is why the negative element responsible for the reversal of polarity transcends from the word-level to the sentence-level rendering the occurrences of almost all negations in Bangla manifest on the syntactic level (Thompson, 2006).
+
+In the cases of double negatives, we see the involvement of lexical negation, a morphological feature that works with negative affixes (prefixes and suffixes) attached to a root word. The prefixes in Bangla have two different phonetic variations or allophones depending on whether the prefix precedes a vowel or a consonant. The same is true for prefixes that imbue a negative connotation to a root word, e.g.  $o$  (ə) and  $on$  (ən). For example,
+
+Kekc 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e 8e
+
+Another negative prefix that precedes a root word to invert its polarity is  $nir$  (NIR). For example,
+
+Leye 1
+
+On the contrary, the suffix  $hin$  (रिन) succeeds a root word to convert it to the corresponding negative form. For example,
+
+Arnrkrn m  n knnnck gne  nna pae aen
+
+The expression of negative sentiment is, therefore, very nuanced in the Bangla language as every occurrence of negative is intertwined with features like the tense, hierarchy of syntax, verb status, case-specific issues, and sequential arrangement of words (Thompson, 2006).
+
+# 6 Conclusion
+
+This paper introduces BANGLABOOK, the largest Bangla book review dataset with 158,065 samples, each labeled with 1 of 3 user sentiments. We provide extensive statistical analysis and strong baselines facilitating the utility of the dataset. Given its massive size and fine-grained sentiment distribution, BANGLABOOK has the potential to alleviate the resource scarcity in Bangla language research.
+
+# 7 Limitations
+
+Many of the reviews that were gathered for constructing BANGLABOOK are discarded because they lack a corresponding rating. A manual annotation process would have yielded a much larger dataset, which was not feasible due to resource constraints. Moreover, one of the challenges for validating the dataset is the lack of statistical models and word-embeddings pre-trained on the Bangla language. Some pre-trained Bangla-BERT models, yet to be trained on extensive corpora, have only recently been proposed. Improving transformer-based models for Bangla can enhance sub-word level contextual understanding which will consequently help in more accurate identification of the sentiments in BANGLABOOK (Islam et al., 2022).
+
+# References
+
+Mst Tuhin Akter, Manoara Begum, and Rashed Mustafa. 2021. Bengali sentiment analysis of e-commerce product reviews using k-nearest neighbors. In 2021 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD), pages 40-44. IEEE.  
+Sanjida Akter and Muhammad Tareq Aziz. 2016. Sentiment analysis on facebook group using lexicon based approach. In 2016 3rd International Conference on Electrical Engineering and Information Communication Technology (ICEEICT), pages 1-4. IEEE.  
+Shad Al Kaiser, Sudipta Mandal, Ashraful Kalam Abid, Ekhfa Hossain, Ferdous Bin Ali, and Intisar Tahmid Naheen. 2021. Social media opinion mining based on bangla public post of facebook. In 2021 24th International Conference on Computer and Information Technology (ICCIT), pages 1-6. IEEE.  
+Abhik Bhattacharjee, Tahmid Hasan, Wasi Ahmad, Kazi Samin Mubasshir, Md Saiful Islam, Anindya Iqbal, M. Sohel Rahman, and Rifat Shahriyar. 2022. BanglaBERT: Language model pretraining and benchmarks for low-resource language understanding evaluation in Bangla. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 1318-1327, Seattle, United States. Association for Computational Linguistics.  
+L. Breiman. 2001. Random forests. Machine Learning, 45:5-32.  
+Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.  
+Shaika Chowdhury and Wasifa Chowdhury. 2014. Performing sentiment analysis in bangla microblog posts. In 2014 International Conference on Informatics, Electronics & Vision (ICIEV), pages 1-6. IEEE.  
+Corinna Cortes and Vladimir Naumovich Vapnik. 1995. Support-vector networks. Machine Learning, 20:273-297.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805.  
+Rajib Chandra Dey and Orvila Sarker. 2019. Sentiment analysis on bengali text using lexicon based approach. In 2019 22nd International Conference on Computer and Information Technology (ICCIT), pages 1-5. IEEE.  
+Asif Hassan, Mohammad Rashedul Amin, Abul Kalam Al Azad, and Nabeel Mohammed. 2016. Sentiment analysis on bangla and romanized bangla text using
+
+deep recurrent models. In 2016 International Workshop on Computational Intelligence (IWCI), pages 51-56. IEEE.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9:1735-1780.  
+MD Iqbal, Avishek Das, Omar Sharif, Mohammed Moshiul Hoque, and Iqbal H Sarker. 2022. Bemoc: A corpus for identifying emotion in bengali texts. SN Computer Science, 3(2):1-17.  
+Khondoker Ittehadul Islam, Md Saiful Islam, and Md Ruhul Amin. 2020. Sentiment analysis in bengali via transfer learning using multi-lingual bert. In 2020 23rd International Conference on Computer and Information Technology (ICCIIT), pages 1-5. IEEE.  
+Khondoker Ittehadul Islam, Sadipta Kar, Md Saiful Islam, and Mohammad Ruhul Amin. 2021. Sentnob: A dataset for analysing sentiment on noisy bangla texts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3265-3271.  
+Khondoker Ittehadul Islam, Tanvir Yuvraz, Md Saiful Islam, and Enamul Hassan. 2022. Emonoba: A dataset for analyzing fine-grained emotions on noisy bangla texts. 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, pages 128-134.  
+George H. John and Pat Langley. 1995. Estimating continuous distributions in bayesian classifiers. In _Conference on Uncertainty in Artificial Intelligence_.  
+Mst Eshita Khatun and Tapasy Rabeya. 2022. A machine learning approach for sentiment analysis of book reviews in bangla language. In 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), pages 1178-1182. IEEE.  
+Saskia le Cessie and J. C. van Houwelingen. 1992. Ridge estimators in logistic regression. Applied statistics, 41:191-201.  
+Tom MY Lin, Pin Luarn, and Yun Kuei Huang. 2005. Effect of internet book reviews on purchase intention: A focus group study. The Journal of Academic Librarianship, 31(5):461-468.  
+Shamsul Arafin Mahtab, Nazmul Islam, and Md Mahfuzur Rahaman. 2018. Sentiment analysis on bangladesh cricket with support vector machine. In 2018 international conference on Bangla speech and language processing (ICBSLP), pages 1-4. IEEE.  
+P Majumder, M Mitra, and BB Chaudhuri. 2002. Ngram: a language independent approach to ir and nlp. In International conference on universal knowledge and language, volume 2.
+
+Muhammad Mahmudun Nabi, Md Tanzir Altaf, and Sabir Ismail. 2016. Detecting sentiment from bangla text using machine learning technique and feature analysis. International Journal of Computer Applications, 153(11):28-34.  
+Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing.  
+Md Rahib, Rumman Hussain Khan, Amzad Hussain Tamim, Mohammad Zawad Tahmeed, and Mohammad Jaber Hossain. 2022. Emotion detection based on bangladeshi peoples social media response on Covid-19. SN Computer Science, 3(2):1-6.  
+Fuad Rahman, Habibur Khan, Zakir Hossain, Mahfuza Begum, Sadia Mahanaz, Ashraful Islam, and Aminul Islam. 2019. An annotated bangla sentiment analysis corpus. In 2019 International Conference on Bangla Speech and Language Processing (ICBSLP), pages 1-5. IEEE.  
+Md Atikur Rahman and Emon Kumar Dey. 2018. Datasets for aspect-based sentiment analysis in bangla and its baseline evaluation. Data, 3(2):15.  
+Gillian Catriona Ramchand. 2004. Two types of negation in bengali. *Clause structure in South Asian languages*, pages 39-66.  
+Sagor Sarker. 2020. Banglabert: Bengali mask language model for bengali language understanding.  
+Salim Sazzed. 2020a. Cross-lingual sentiment classification in low-resource bengali language. In Proceedings of the sixth workshop on noisy user-generated text (W-NUT 2020), pages 50–60.  
+Salim Sazzed. 2020b. Development of sentiment lexicon in bengali utilizing corpus and cross-lingual resources. In 2020 IEEE 21st International conference on information reuse and integration for data science (IRI), pages 237-244. IEEE.  
+Salim Sazzed. 2021. Bengsentilex and bengswear-lex: creating lexicons for sentiment analysis and profanity detection in low-resource bengali language. PeerJ Computer Science, 7:e681.  
+Marina Sokolova, Nathalie Japkowicz, and S. Szpakowicz. 2006. Beyond accuracy, f-score and roc: A family of discriminant measures for performance evaluation. In Australian Conference on Artificial Intelligence.  
+Alan T Sorensen and Scott J Rasmussen. 2004. Is any publicity good publicity? a note on the impact of book reviews. NBER Working paper, Stanford University.  
+Nusrath Tabassum and Muhammad Ibrahim Khan. 2019. Design an empirical framework for sentiment analysis from bangla text using machine learning. In 2019 International Conference on Electrical, Computer and Communication Engineering (ECCE), pages 1-5. IEEE.
+
+SM Abu Taher, Kazi Afsana Akhter, and KM Azharul Hasan. 2018. N-gram based sentiment mining for bangla text using support vector machine. In 2018 international conference on Bangla speech and language processing (ICBSLP), pages 1-5. IEEE.  
+Hanne-Ruth Thompson. 2006. Negation patterns in bengali. Bulletin of the School of Oriental and African Studies, 69(2):243-265.  
+Rashedul Amin Tuhin, Bechitra Kumar Paul, Faria Nawrine, Mahbuba Akter, and Amit Kumar Das. 2019. An automated system of sentiment analysis from bangla text using supervised learning techniques. In 2019 IEEE 4th International Conference on Computer and Communication Systems (ICCCS), pages 360-364. IEEE.  
+Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. ArXiv, abs/1706.03762.  
+Anning Wang, Qiang Zhang, Shuangyao Zhao, Xiaonong Lu, and Zhanglin Peng. 2020. A review-driven customer preference measurement model for product improvement: sentiment-based importance performance analysis. Information Systems and e-Business Management, 18:61-88.  
+Alice Zheng and Amanda Casari. 2018. Feature engineering for machine learning: principles and techniques for data scientists. "O'Reilly Media, Inc."
+
+A Appendix
+
+<table><tr><td>Dataset</td><td colspan="2">Sentiment Classification</td><td colspan="2">Sentiment Distribution</td><td>Total # of Samples</td><td>Availability</td><td>Type of Content</td><td>Source(s)</td><td>Baseline Models</td></tr><tr><td>(Tabassum and Khan, 2019)</td><td colspan="2">Positive Negative</td><td colspan="2">- -</td><td>1,050</td><td>Closed</td><td>Posts, comments</td><td>Facebook, Twitter</td><td>RF</td></tr><tr><td>(Chowdhury and Chowdhury, 2014)</td><td colspan="2">Positive Negative</td><td colspan="2">- -</td><td>1,300</td><td>Closed</td><td>Posts, comments</td><td>Twitter</td><td>-</td></tr><tr><td>(Nabi et al., 2016)</td><td colspan="2">Positive Negative</td><td colspan="2">- -</td><td>1,500</td><td>Closed</td><td>Posts, comments</td><td>Social Media</td><td>-</td></tr><tr><td>(Mahtab et al., 2018)</td><td colspan="2">Praise Criticism Sadness</td><td colspan="2">513 604 484</td><td>1,601</td><td>Closed</td><td>Comments</td><td>Prothom Alo Online News Portal</td><td>SVM, DT, NB</td></tr><tr><td>(Akter and Aziz, 2016)</td><td colspan="2">Positive Negative Neutral</td><td colspan="2">- -</td><td>3,600</td><td>Closed</td><td>Posts, comments</td><td>Facebook</td><td>NB</td></tr><tr><td>(Rahman and Kumar Dey, 2018)</td><td colspan="2">Positive Negative Neutral</td><td colspan="2">- -</td><td>4,700</td><td>Open</td><td>Comments</td><td>Facebook pages: BBC Bangla, Prothom Alo</td><td>SVM, LR, KNN, DT, LSTM NB, CNN</td></tr><tr><td>(Dey and Sarker, 2019)</td><td colspan="2">Positive Negative</td><td colspan="2">2,600 2,600</td><td>5,200</td><td>Closed</td><td>Comments, reviews</td><td>Facebook, Twitter, YouTube, News Portals</td><td>DT, NB, SVM</td></tr><tr><td>(Khatun and Rabeya, 2022)</td><td colspan="2">Positive Negative</td><td colspan="2">- -</td><td>5,500</td><td>Closed</td><td>Comments, reviews</td><td>Social Media</td><td>-</td></tr><tr><td>BEMOC (Iqbal et al., 2022)</td><td colspan="2">AngerFearSadJoyDisgust</td><td colspan="2">- -</td><td>7,000</td><td>Open</td><td>Posts, comments</td><td>Facebook, YouTube, Online blogs, Bangla story books, novels, newspapers, discourse</td><td>-</td></tr><tr><td>(Tuhin et al., 2019)</td><td colspan="2">Happy Tender Excited Sad Angry Scared</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>(Akter et al., 2021)</td><td colspan="2">Positive Negative Neutral</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td>7,905</td><td>Closed</td><td>Product reviews</td><td>Daraz</td><td>RF, LR, SVM, KNN, XGB</td><td></td></tr><tr><td>(Hassan et al., 2016)</td><td colspan="2">Positive Negative Ambiguous</td><td>- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>(Rahib et al., 2022)</td><td colspan="2">Insightful Curious Gratitude</td><td colspan="2">3,800 3,549 3,232</td><td>10,581</td><td>Closed</td><td>Comments</td><td>Social Media</td><td>SVM, RF, CNN, LSTM</td></tr><tr><td rowspan="2">(Al Kaiser et al., 2021)</td><td rowspan="2">Positive Negative Neutral</td><td>Wishful Thinking Appreciation</td><td>967 942</td><td>1,909</td><td rowspan="2">11,006</td><td rowspan="2">Closed</td><td rowspan="2">Comments</td><td rowspan="2">Facebook</td><td rowspan="2">LR, DT, RF, MNB, KNN, Linear SVM, RBF SVM, XGB</td></tr><tr><td>Gender-based hate Religious hate Political hate Personal hate Sarcasm</td><td>525 731 572 1,995 1,414 3,860</td><td>5,237 3,860</td></tr><tr><td>(Sazzed, 2020a)</td><td colspan="2">Positive Negative</td><td colspan="2">8,500 3,307</td><td>11,807</td><td>Open</td><td>Comments</td><td>YouTube</td><td>SVM, ET, RF, LR, VADER, TextBlob</td></tr><tr><td>(Sazzed, 2020b)</td><td colspan="2">Positive Negative</td><td colspan="2">- -</td><td>12,000</td><td>Closed</td><td>Comments</td><td>YouTube</td><td>-</td></tr><tr><td rowspan="2">SENTNoB (Islam et al., 2021)</td><td rowspan="2" colspan="2">Positive Negative</td><td rowspan="2" colspan="2">6,410 5,709 3,609</td><td rowspan="2">15,728</td><td rowspan="2">Open</td><td rowspan="2">Comments</td><td rowspan="2">Prothom Alo Online Newspaper, YouTube</td><td rowspan="2">RNN</td></tr><tr></tr><tr><td>(Islam et al., 2020)</td><td colspan="2">Positive Negative Neutral</td><td colspan="2">4,769 8,351 4,732</td><td>17,852</td><td>Open</td><td>Comments</td><td>Prothom Alo Online Newspaper</td><td>CNN, LSTM, BERT, GRU, fastText</td></tr><tr><td>EMONOBA (Islam et al., 2022)</td><td colspan="2">Love Joy Surprise Anger Sadness Fear</td><td colspan="2">4,202 9,249 939 3,905 5,109 307</td><td>20,468</td><td>Open</td><td>Comments</td><td>YouTube, Facebook, Twitter, Prothom Alo</td><td>Bi-LSTM, fastText, Bangla-BERT-base</td></tr><tr><td>BANGLABOOK (ours)</td><td colspan="2">Positive Negative Neutral</td><td colspan="2">141,587 9,674 6,804</td><td>158,065</td><td>\( Open^† \)</td><td>Book reviews</td><td>Rokomari, Wafilife</td><td>RF, LSTM, LR, GRU MNB, SVM, XGB Bangla-BERT</td></tr></table>
+
+Table 5: Comparison of notable Bangla Sentiment Analysis datasets sorted in ascending order of size. The abbreviations respectively denote, Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), Naive Bayes (NB), Logistic Regression (LR), K-Nearest Neighbors (KNN), Long Short-term Memory (LSTM), Convolutional Neural Network (CNN), Extreme Gradient Boost (XGB), Multinomial Naive Bayes (MNB), Radial Basis Function (RBF), Extreme Random Tree (ET), Recurrent Neural Network (RNN), Bidirectional Encoder Representations from Transformers (BERT), Gated Recurrent Unit (GRU), Bi-LSTM (Bidirectional LSTM). All the publicly available datasets are hyperlinked. Open† denotes the redaction of the link for anonymity.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Left blank.
+
+A2. Did you discuss any potential risks of your work?
+
+Not applicable. Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+Not applicable. Left blank.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Not applicable. Left blank.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Not applicable. Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Not applicable. Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Not applicable. Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Left blank.
+
+C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Left blank.
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank.

2023/BanglaBook_ A Large-scale Bangla Dataset for Sentiment Analysis from Book Reviews/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:363792c1d34d82035d241a58b36d86ccc119adf4e86fd3ad4d6976a0f75480c4
+size 470369