claim_id large_stringlengths 20 20 | paper_id large_stringlengths 8 36 | year int32 1.96k 2.03k | atomic_claim large_stringlengths 17 500 | claim_index int32 0 12 | extractor_model large_stringclasses 1
value | prompt_hash large_stringclasses 1
value | run_id large_stringclasses 425
values | created_at large_stringclasses 425
values |
|---|---|---|---|---|---|---|---|---|
clm_08cebd97fdc03047 | 2020.emnlp-main.119 | 2,020 | Analogous Process Structure Induction (APSI) framework leverages analogies among processes and conceptualization of sub-event instances to predict whole sub-event sequences of previously unseen open-domain processes. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_9f8046117da2e278 | 2020.emnlp-main.119 | 2,020 | APSI generates meaningful sub-event sequences for unseen processes. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_9642b5f4dcc5ce39 | 2020.emnlp-main.119 | 2,020 | APSI can predict missing events in sub-event sequences. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_5361a82ebc93ccd9 | 2020.emnlp-main.107 | 2,020 | A neural-based NER approach for social media texts incorporates both local context and augmented semantic information. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_78bbcf1c7025b59a | 2020.emnlp-main.107 | 2,020 | Augmented semantic information is derived from a large-scale corpus. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_15d63cb9a902b442 | 2020.emnlp-main.107 | 2,020 | An attentive semantic augmentation module encodes the augmented semantic information. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_bfce31461bc3df7a | 2020.emnlp-main.107 | 2,020 | A gate module aggregates local and augmented semantic representations. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_0643bc14b5bb2afb | 2020.emnlp-main.107 | 2,020 | The approach outperforms prior methods on three benchmark social media NER datasets in English and Chinese. | 4 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_a7907e5d359672f1 | 2020.emnlp-main.54 | 2,020 | Generation with Multi-Hop Reasoning Flow (GRF) enables pre-trained language models to perform dynamic multi-hop reasoning over multi-relational paths from an external commonsense knowledge graph. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_56acbe3d0de76ff9 | 2020.emnlp-main.54 | 2,020 | GRF outperforms existing baselines on three text generation tasks that require reasoning over commonsense knowledge. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_7ff999b68e147601 | 2020.emnlp-main.54 | 2,020 | The dynamic multi-hop reasoning module generates reasoning paths that provide rationales for the generated text. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_6dcd2f1d0a725413 | 2020.emnlp-main.76 | 2,020 | Token-level adaptive objectives based on token frequencies assign larger weights to low-frequency target tokens during NMT training. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_656ae05f3bd57495 | 2020.emnlp-main.76 | 2,020 | The adaptive training method improves translation quality across multiple language pairs. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_8eb7c807e2d03104 | 2020.emnlp-main.76 | 2,020 | The adaptive training method yields larger translation quality gains on sentences that contain more low-frequency tokens. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1761407ce6d4fc83 | 2020.emnlp-main.76 | 2,020 | The adaptive training method improves lexical diversity of the generated translations. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_22e29dc08cf51cd9 | 2020.emnlp-main.68 | 2,020 | A scoring method evaluates utterance pair quality based on connectivity and content relatedness. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_73540cd8a4e69016 | 2020.emnlp-main.68 | 2,020 | The scoring method shows relatively good correlation with human judgments of dialogue quality. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_52c72cbac9acc34f | 2020.emnlp-main.68 | 2,020 | The scoring method can be used to filter potentially unacceptable utterance pairs from large-scale noisy dialogue corpora. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_96787b66f8fbce76 | 2020.emnlp-main.68 | 2,020 | Filtering training data with the scoring method improves response generation quality of neural dialogue agents. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_f8362ce390e0b39b | 2020.emnlp-main.52 | 2,020 | Knowledge Consolidation Network (KCN) enables incremental event detection without performance degradation on previously learned event types. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_eb459a258d24922c | 2020.emnlp-main.52 | 2,020 | Prototype enhanced retrospection mitigates semantic ambiguity between old and new event classes in incremental event detection. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_078ee04beb69be9b | 2020.emnlp-main.52 | 2,020 | Hierarchical distillation mitigates class imbalance between old and new event classes in incremental event detection. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_e1c52720e1990ea7 | 2020.emnlp-main.52 | 2,020 | KCN achieves higher whole F1 scores than prior state-of-the-art models on the ACE and TAC KBP benchmarks. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_6211731fe0e0f6fe | 2020.emnlp-main.125 | 2,020 | An unordered tree mapping algorithm combined with phrase representation modelling that embeds similarity distributions onto contextualized representations addresses phrase alignment. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_4933fb0873e7bc3e | 2020.emnlp-main.125 | 2,020 | The approach effectively handles both compositional and non-compositional global phrase alignments. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_eda7bc5416203c04 | 2020.emnlp-main.125 | 2,020 | The approach outperforms the method used in a previous study. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_b0f2936e2ee640f0 | 2020.emnlp-main.125 | 2,020 | The approach achieves performance competitive with experienced human annotators. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_e8a7f4016739563b | 2020.emnlp-main.87 | 2,020 | Iterative generalizations of the roundtrip consistency method are introduced to maximize an approximation of a lower bound on the probability of adaptation data. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_c08eb35002380485 | 2020.emnlp-main.87 | 2,020 | Unsupervised self-adaptation of question answering systems is achieved by iteratively generating synthetic QA pairs on target data. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_e1851b2441e0fe87 | 2020.emnlp-main.87 | 2,020 | The proposed method improves question answering performance while requiring substantially less synthetic data and training computation than existing augmentation approaches. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_632e9892f3080030 | 2020.emnlp-main.80 | 2,020 | Self-paced learning enables neural machine translation models to automatically quantify learning confidence over training examples. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_31858e45a7a74410 | 2020.emnlp-main.80 | 2,020 | Self-paced learning enables neural machine translation models to flexibly govern learning by regulating loss at each iteration. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_c7f881bbc3821a84 | 2020.emnlp-main.80 | 2,020 | Self-paced learning improves translation quality compared to strong baselines and models trained with human-designed curricula. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_d04e57adfc00e056 | 2020.emnlp-main.80 | 2,020 | Self-paced learning accelerates convergence compared to strong baselines and models trained with human-designed curricula. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1223ca5a1246306b | 2020.emnlp-main.108 | 2,020 | A Hierarchical Transformer divides long social media threads into shorter subthreads, represents each subthread with BERT, and encodes all subthreads with a global Transformer to obtain thread representations. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_595153da3d6b29a4 | 2020.emnlp-main.108 | 2,020 | A Coupled Transformer Module captures inter-task interactions between stance classification and rumor verification. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1929780bd39edb96 | 2020.emnlp-main.108 | 2,020 | A Post-Level Attention layer incorporates predicted stance labels to improve rumor verification. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_90bd100b45f1735a | 2020.emnlp-main.108 | 2,020 | The Coupled Hierarchical Transformer model outperforms existing multi-task learning approaches on two benchmark rumor verification datasets. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_031eb4fe483ccec4 | 2020.emnlp-main.64 | 2,020 | Debiased-Chat is an adversarial learning framework that trains dialogue models to be free from gender bias while preserving performance. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_242c423981a4453b | 2020.emnlp-main.64 | 2,020 | Debiased-Chat reduces gender bias in dialogue models and maintains response quality on two real-world conversation datasets. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_8bb1b3cd6143944f | 2020.emnlp-main.78 | 2,020 | Pruned parameters can be rejuvenated to improve baseline neural machine translation performance. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_2e665ad6a26341d0 | 2020.emnlp-main.78 | 2,020 | Rejuvenated parameters are reallocated to enhance the modeling of low-level lexical information. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_712c93bf9bbe3618 | 2020.emnlp-main.78 | 2,020 | Systematic experiments and analyses are conducted on multiple datasets and neural machine translation architectures to assess the reuse of redundant parameters. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_6e7473a86616057c | 2020.emnlp-main.111 | 2,020 | A large-scale medical multiple-choice question dataset for the National Licensed Pharmacist Examination in China is collected. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_c735634df08e55e8 | 2020.emnlp-main.111 | 2,020 | KMQA is a reading comprehension model that fully exploits structural medical knowledge from a knowledge graph and reference medical plain text retrieved from books. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_a4f9c444d4747e15 | 2020.emnlp-main.111 | 2,020 | KMQA outperforms existing competitive models by a large margin and achieves exam-passing level accuracy on the test set. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_f60ab1673fcb7a3d | 2020.emnlp-main.13 | 2,020 | A cross-formalism layer probing study in role semantics was conducted. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_b45dd9fd39939d2f | 2020.emnlp-main.13 | 2,020 | BERT encodes semantic role and proto-role information differently depending on the linguistic formalism used. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_3880989ba688169e | 2020.emnlp-main.13 | 2,020 | Layer probing can detect subtle differences between implementations of the same linguistic formalism. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_a1987d91ed38303c | 2020.emnlp-main.13 | 2,020 | Linguistic formalism constitutes an important dimension in probing studies alongside cross-task and cross-lingual settings. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_6a8f4cfe167e29a5 | 2020.emnlp-main.109 | 2,020 | A novel prediction task of attribution tie detection is defined for identifying factors held responsible for a crisis in social media documents. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_603584e600441de8 | 2020.emnlp-main.109 | 2,020 | A neural baseline model is presented for identifying attribution ties in social media comments. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_14b03a3b43fd8d4f | 2020.emnlp-main.109 | 2,020 | The first annotated dataset of 2,500 YouTube comments about the 2019 Chennai water crisis is released. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_be5a781aff2fd8a4 | 2020.emnlp-main.81 | 2,020 | Long-short term masking self-attention captures long-range dependence in document-level neural machine translation. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_22e5abeec30b6325 | 2020.emnlp-main.81 | 2,020 | Long-short term masking self-attention reduces error propagation in document-level neural machine translation. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_913ba99528ef470d | 2020.emnlp-main.81 | 2,020 | The proposed method achieves strong translation quality on publicly available document-level datasets. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_b5d389892c2f905e | 2020.emnlp-main.81 | 2,020 | The proposed method captures discourse phenomena in document-level translation. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_66274daa67ac3503 | 2020.emnlp-main.81 | 2,020 | An extensive analysis of the standard transformer shows that its auto-regressive property provides consistency but also leads to error accumulation in document-level translation. | 4 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_a8377c513734684d | 2020.emnlp-main.90 | 2,020 | A large-scale dataset for KG-to-text generation that includes abundant knowledge of various entity types and reflects the practical scenario where descriptions cover only the most significant knowledge is introduced. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1e6209111c3a7e78 | 2020.emnlp-main.90 | 2,020 | A multi-graph structure that represents original graph information more comprehensively is proposed for KG-to-text generation. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_ec930f51d92e48f3 | 2020.emnlp-main.90 | 2,020 | Aggregation methods that learn to extract rich graph information are incorporated into the model architecture. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_a5845fb9e4e6e71f | 2020.emnlp-main.90 | 2,020 | Experiments demonstrate that the proposed model architecture is effective for generating entity descriptions from knowledge graphs. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_8eed49bccb30fdb2 | 2020.emnlp-main.1 | 2,020 | A large-scale analysis of sentence attackability in online arguments is presented. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_c0d3ba0ab907ac72 | 2020.emnlp-main.1 | 2,020 | Driving reasons for attacks in argumentation are analyzed and relevant sentence characteristics are identified. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1139bd272cc5b68b | 2020.emnlp-main.1 | 2,020 | Sentence attackability is associated with characteristics of sentence content, proposition types, and tone. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_5e336121f4ed15b3 | 2020.emnlp-main.1 | 2,020 | An external knowledge source provides useful information for predicting sentence attackability. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_0dffc0a8b4f5c5d6 | 2020.emnlp-main.1 | 2,020 | Machine learning models can automatically detect attackable sentences in arguments, achieving performance comparable to laypeople and surpassing several baselines. | 4 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_810963e71310aabd | 2020.emnlp-main.24 | 2,020 | FIND enables humans to debug deep learning text classifiers by disabling irrelevant hidden features. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_5a6c517955b782ad | 2020.emnlp-main.24 | 2,020 | Using FIND, humans improve CNN text classifiers trained on biased datasets. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_b275820d2af9687c | 2020.emnlp-main.24 | 2,020 | Using FIND, humans improve CNN text classifiers trained on datasets with dissimilar train-test distributions. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_e3e46e3321e79851 | 2020.emnlp-main.92 | 2,020 | Back-predicting projected AMR graphs on the target sentence during generation serves as a decoder for AMR-to-text generation. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_17c493d0f528b704 | 2020.emnlp-main.92 | 2,020 | The back-prediction decoder generates text that better preserves the input AMR meaning than standard language-modeling decoders. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_b49dcb19799af536 | 2020.emnlp-main.92 | 2,020 | The back-prediction decoder achieves superior performance compared to the prior graph Transformer state-of-the-art on two AMR-to-text benchmarks. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_bf2d11d995d080d6 | 2020.emnlp-main.11 | 2,020 | Knowledge Triplet Learning is a self-supervised task over knowledge graphs. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1dd249c3f256bc05 | 2020.emnlp-main.11 | 2,020 | Heuristics are introduced to create synthetic graphs for commonsense and scientific knowledge. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_7e1f67071da1861c | 2020.emnlp-main.11 | 2,020 | Knowledge Triplet Learning enables zero-shot question answering. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_0a41c8b2000e43ae | 2020.emnlp-main.11 | 2,020 | Knowledge Triplet Learning yields considerable improvements over large pre-trained transformer language models on zero-shot question answering. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_640279c675bc287e | 2020.emnlp-main.7 | 2,020 | Simulated Multiple Reference Training (SMRT) approximates the full space of possible translations by sampling paraphrases of reference sentences and training the MT model to predict the paraphraser’s token distribution. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1b614f76a9ba1274 | 2020.emnlp-main.7 | 2,020 | SMRT improves translation quality in low-resource machine translation when translating to English. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_a03f9c5a85f6eb15 | 2020.emnlp-main.7 | 2,020 | SMRT is complementary to back-translation in low-resource machine translation. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_3722559c97cacd32 | 2020.emnlp-main.19 | 2,020 | Extended Transformer Construction (ETC) is a Transformer architecture that scales input length and encodes structured inputs. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_9708ba32d56b998e | 2020.emnlp-main.19 | 2,020 | ETC introduces a global-local attention mechanism between global tokens and regular input tokens. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_8aa8bcab710757dd | 2020.emnlp-main.19 | 2,020 | ETC combines global-local attention with relative position encodings and a Contrastive Predictive Coding pre-training objective to encode structured inputs. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_1b23684ec08b27f4 | 2020.emnlp-main.19 | 2,020 | ETC achieves state-of-the-art performance on four natural language datasets that require long and/or structured inputs. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_73ad8f5d81d65de6 | 2020.emnlp-main.127 | 2,020 | Graph Aggregation-and-Inference Network (GAIN) is a method for document-level relation extraction that recognizes relations among entities in long paragraphs. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_265867e0d807013d | 2020.emnlp-main.127 | 2,020 | GAIN constructs a heterogeneous mention-level graph that captures complex interactions among different mentions. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_e29e2ea01e5d45ad | 2020.emnlp-main.127 | 2,020 | GAIN constructs an entity-level graph that aggregates mentions belonging to the same entity. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_51f4c30155b7b900 | 2020.emnlp-main.127 | 2,020 | GAIN employs a novel path reasoning mechanism over the constructed graphs to infer relations between entities. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_30ff91285d37e46d | 2020.emnlp-main.127 | 2,020 | GAIN achieves performance improvement over previous state-of-the-art methods on the DocRED dataset. | 4 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_5382b2a16135cf2d | 2020.emnlp-main.49 | 2,020 | Event extraction can be performed by formulating it as a question answering task that extracts event arguments end-to-end. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_c764a59635ae4945 | 2020.emnlp-main.49 | 2,020 | A question answering based event extraction framework avoids error propagation from entity recognition preprocessing. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_2d29522d50b34329 | 2020.emnlp-main.49 | 2,020 | The question answering based event extraction framework outperforms prior event extraction methods. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_8c16b6128b6269bf | 2020.emnlp-main.49 | 2,020 | The question answering based event extraction framework can extract event arguments for roles not seen during training (zero-shot learning). | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_ff35d5e2edcc6b14 | 2020.emnlp-main.70 | 2,020 | An evaluation framework based on psycholinguistic referential bias contexts is used to assess discourse modeling ability of neural language models. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_730d69bdab52a649 | 2020.emnlp-main.70 | 2,020 | Neural language models' prediction behavior of reference largely diverges from human referential biases. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_dabf4714f5901678 | 2020.emnlp-main.36 | 2,020 | Contrastive Distillation on Intermediate Representations (CoDIR) is a knowledge distillation framework that trains a student model to distill knowledge through intermediate layers of a teacher via a contrastive objective. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_de3572acd15ba6c2 | 2020.emnlp-main.36 | 2,020 | CoDIR enables the student model to exploit rich information in the teacher’s hidden layers by distinguishing a positive sample from many negative samples. | 1 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_3cde1f08039f6085 | 2020.emnlp-main.36 | 2,020 | CoDIR can be applied to compress large-scale language models during both pre-training and fine-tuning stages. | 2 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_9f6ad9708120dc47 | 2020.emnlp-main.36 | 2,020 | CoDIR achieves superior performance on the GLUE benchmark compared to prior state-of-the-art compression methods. | 3 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
clm_9bab20c0122fad53 | 2020.emnlp-main.126 | 2,020 | Structure-Aware Transformer injects table structural information into the self-attention mask. | 0 | Openai/Gpt-oss-120b | f9a9b434c72c | extract_20260701T234408Z | 2026-07-01T23:44:08.237282+00:00 |
ACL Anthology Atomic Contribution Claims (ACC)
346,010 atomic contribution claims extracted from the abstracts of 80,144 ACL Anthology papers — 423 venues, 1963–2026.
An atomic contribution claim (ACC) is a single self-contained sentence stating one concrete contribution of a paper: atomic (exactly one contribution-bearing proposition), decontextualized (pronouns resolved, meta-language removed), and falsifiable (a verifiable assertion). Unlike raw abstracts or keywords, ACCs separate what a paper contributes from what it merely mentions — the word "parsing" in the motivation of an LLM paper never becomes a parsing claim.
Analogous Process Structure Induction (APSI) framework leverages analogies among processes and conceptualization of sub-event instances to predict whole sub-event sequences of previously unseen open-domain processes. — from 2020.emnlp-main.119
This is the full-anthology corpus of the Drift Inspector project:
- 🔭 *Live demo (extended, 6 ACL venues): https://hamyrappy.github.io/drift-inspector-acl/
- 📄 EMNLP paper case study instance: https://hamyrappy.github.io/drift-inspector-emnlp/
- 🛠 Code (extraction + clustering + inspector): https://github.com/Hamyrappy/drift-inspector
- 🤗 Companion dataset (EMNLP 2020–2025, with drift clusters): emnlp-2020-2025-atomic-claims
What's inside
| Config | Rows | Contents |
|---|---|---|
claims (default) |
346,010 | one row per extracted atomic claim |
papers |
80,144 | bibliographic metadata of every source paper |
Coverage: every ACL Anthology paper with an abstract available through the acl-anthology library at extraction time (80,144 of 125,354 anthology papers; abstract availability is much higher from 2018 onward). 499 abstracts yielded no extractable contribution (0 claims). Mean 4.3 claims per paper.
Top venues by claims: findings 57,699 · acl 50,268 · emnlp 40,999 · lrec 31,010 · naacl 15,825 · coling 12,463 · semeval 11,056. Claims are written in English even when the source abstract is not (e.g. TALN, CCL, ROCLING).
claims fields
| Column | Type | Description |
|---|---|---|
claim_id |
string | unique id (clm_ + hash) |
paper_id |
string | ACL Anthology id, e.g. 2020.emnlp-main.119 — joins to papers |
year |
int32 | publication year of the source paper |
atomic_claim |
string | the claim text |
claim_index |
int32 | position of the claim within its paper's extraction (0-based) |
extractor_model |
string | Openai/Gpt-oss-120b |
prompt_hash |
string | hash of the exact extraction prompt (f9a9b434c72c; full prompt in the code repo) |
run_id, created_at |
string | extraction run provenance |
papers fields
| Column | Type | Description |
|---|---|---|
paper_id |
string | ACL Anthology id |
year |
int32 | publication year (1963–2026; 2026 is a partial year) |
title |
string | paper title |
venue |
string | primary anthology venue id (e.g. acl, findings, wmt) |
url |
string | https://aclanthology.org/<paper_id>/ |
authors |
string | author names, ;-separated |
author_ids |
string | anthology author ids, ;-separated |
n_authors |
int32 | number of authors |
Abstracts are not redistributed here; re-fetch them by paper_id via the acl-anthology Python library if needed.
from datasets import load_dataset
claims = load_dataset("Hamyrappy/acl-anthology-atomic-claims", "claims", split="train")
papers = load_dataset("Hamyrappy/acl-anthology-atomic-claims", "papers", split="train")
How this differs from existing ACL Anthology corpora
Corpora such as ACL-OCL or the ACL Anthology Corpus redistribute paper text (abstracts/full text) and metadata. This dataset instead provides a derived semantic unit: each abstract is decomposed into individually usable contribution statements, suitable for claim-level clustering, retrieval, drift/trend analysis, and science-of-science studies where "what did this paper actually contribute?" is the unit of interest.
Dataset creation
- Corpus. All ACL Anthology papers with abstracts, enumerated via the official
acl-anthologylibrary (metadata snapshot: July 2026). - Extraction.
openai/gpt-oss-120b(reasoning effort medium), few-shot prompt enforcing atomicity / decontextualization / falsifiability and excluding background, motivation, related work, and raw metric-only claims. One LLM pass per abstract, temperature 0.2. Zero extraction errors across the corpus; the exact prompt (hashf9a9b434c72c) ships with the code. - Quality monitoring. An independent LLM judge (
gemini-2.5-flash-lite, separately prompted) rates a stratified 300-claim sample 96.7% Good. The judge protocol was itself calibrated on a human-validated EMNLP subset (three annotators, Fleiss' κ = 0.844; judge–human agreement 94.5%, κ = 0.863) — see the companion dataset for the human-validation corpus and protocol.
Limitations
- LLM-extracted, not human-curated. Claims can contain extraction errors (hallucinated specifics, under-/over-splitting, background leakage). Quality is judge-monitored on samples, not human-revalidated per row.
- Abstract-level. Contributions that surface only in the paper body are invisible; abstracts also omit negative results and technical nuance.
- Coverage is abstract-availability-biased. Pre-2000s papers rarely have machine-readable abstracts; per-year counts reflect anthology coverage, not publication volume.
- Claim counts per paper vary with abstract style and venue norms; treat cross-venue frequency comparisons descriptively.
Licensing
The claim texts are LLM-generated derived statements and are released under CC BY 4.0, matching the ACL Anthology's license for materials published in or after 2016 (which cover 92.6% of source papers here). Bibliographic metadata (titles, authors, years, URLs) are factual records. Note that the ACL Anthology licenses materials prior to 2016 under CC BY-NC-SA 3.0; claims derived from pre-2016 abstracts (6.7% of rows) can be excluded by filtering year >= 2016 if your use requires it. Source abstracts are not redistributed in this dataset. Please attribute the ACL Anthology and cite the underlying papers where appropriate.
Citation
The system paper introducing ACCs and Drift Inspector is under review; until then, please cite the dataset directly:
@misc{acc-acl-anthology-2026,
title = {ACL Anthology Atomic Contribution Claims (ACC)},
author = {Karimov, Vsevolod and Ostarkov, Stepan and Poroshina, Anastasia and Frolov, Anatoly and Panchenko, Alexander},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/Hamyrappy/acl-anthology-atomic-claims}}
}
- Downloads last month
- 53