claim_id
large_stringlengths
20
20
paper_id
large_stringlengths
17
20
year
int32
2.02k
2.03k
atomic_claim
large_stringlengths
18
387
claim_index
int32
0
8
extractor_model
large_stringclasses
1 value
prompt_hash
large_stringclasses
1 value
run_id
large_stringclasses
2 values
created_at
large_stringclasses
2 values
clm_0ea6495e0a53d81c
2021.emnlp-main.602
2,021
SRU++ is a highly-efficient architecture that combines fast recurrence and attention for sequence modeling.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_b60a1755be95a7da
2021.emnlp-main.602
2,021
SRU++ achieves better bits-per-character and perplexity with reduced training cost compared to top-performing Transformer models on standard language modeling tasks.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_639ae64d6fc7d8c2
2021.emnlp-main.602
2,021
SRU++ requires minimal attention to achieve near state-of-the-art performance.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_7979ce943bfc1a36
2020.emnlp-main.593
2,020
DyERNIE learns evolving entity representations in a product of Riemannian manifolds, with the composed spaces estimated from the sectional curvatures of the data.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_915778e463e97925
2020.emnlp-main.593
2,020
DyERNIE uses a velocity vector defined in the tangent space at each timestamp to capture the evolutionary dynamics of temporal knowledge graphs.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0c43f723fe783c91
2020.emnlp-main.593
2,020
The dynamics of multi-relational graph data can be more properly modeled by the evolution of embeddings on Riemannian manifolds.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_b71b970484e2e00c
2023.emnlp-main.50
2,023
EMMA is a unified encoder-decoder model that reasons over images and trajectories and casts action prediction as multimodal text generation.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ba82e4c6f9eb7a25
2023.emnlp-main.50
2,023
EMMA learns a language of actions by unifying tasks as text generation, which facilitates transfer across tasks.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_e648172b2ded01ac
2023.emnlp-main.50
2,023
EMMA is a single multitask model where each task contributes to goal completion.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0188bd1415d63a71
2023.emnlp-main.50
2,023
EMMA sets a new state-of-the-art on the Dialog-guided Task Completion benchmark.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_128c52b58d598081
2025.emnlp-main.927
2,025
MultiLogicNMR is a multi-extension non-monotonic reasoning dataset that includes one variant with more extensions and one variant with text diversity.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ae87cabdefb99ee5
2025.emnlp-main.927
2,025
MultiLogicNMRer is a neural-symbolic framework for multi-extension non-monotonic reasoning.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_a86e5380f95615af
2025.emnlp-main.927
2,025
Large language models face significant challenges in non-monotonic reasoning with multiple extensions.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_462e4aeef291d122
2025.emnlp-main.927
2,025
MultiLogicNMRer achieves higher accuracy for multi-extension non-monotonic reasoning than prompt-based methods and some fine-tuning methods.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_31597da66bc09eda
2020.emnlp-main.490
2,020
Adversarial Semantic Decoupling is a robust and model-agnostic slot filling method that decouples local semantics inherent in open-vocabulary slot words from the global context.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ed6d6dba855c72e7
2020.emnlp-main.490
2,020
Adversarial Semantic Decoupling outperforms other methods on open-vocabulary slots while maintaining performance on normal slots.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0bb226ddef92ac90
2020.emnlp-main.747
2,020
Human rationales do not necessarily perform well on automatic rationale evaluation metrics.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0823c041fc3a78e4
2020.emnlp-main.747
2,020
Improved rationale evaluation metrics account for model-dependent baseline performance.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_5eedced2c55be837
2020.emnlp-main.747
2,020
Model retraining characterizes rationale quality.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_46c36168027b8a5a
2020.emnlp-main.747
2,020
Fidelity curves reveal properties of rationale quality such as irrelevance and redundancy.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_2bc118f478d2fc61
2022.emnlp-main.341
2,022
MetaFill is a text-infilling-based approach for meta-path generation.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ac863a678e2f871d
2022.emnlp-main.341
2,022
MetaFill formulates meta-path identification as a word sequence infilling problem solvable by pretrained language models.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_c1044d9e357758ea
2022.emnlp-main.341
2,022
MetaFill outperforms existing meta-path generation methods and graph embedding methods that do not leverage meta-paths in link prediction and node classification.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0cd7845c1433c620
2022.emnlp-main.341
2,022
MetaFill accurately classifies edges in the zero-shot setting where existing approaches fail to generate meta-paths.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_aab433241b71f4f4
2023.emnlp-main.862
2,023
AfriSenti is a Twitter sentiment analysis benchmark for 14 African languages from four language families, annotated by native speakers.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_42b99731d5fb4b16
2024.emnlp-main.12
2,024
NumeroLogic represents numbers by prefixing the digit count (e.g., '2:42').
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_11333fa71a4ca50f
2024.emnlp-main.12
2,024
NumeroLogic serves as a Chain of Thought for number generation by requiring the model to consider the digit count first.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_1eeba5562cc51874
2024.emnlp-main.12
2,024
NumeroLogic improves numerical reasoning on arithmetic tasks.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_2e2f40f72887c4d1
2024.emnlp-main.12
2,024
NumeroLogic improves language understanding on the MMLU benchmark.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_cbde504188c59d44
2020.emnlp-main.320
2,020
Training regularization with propositional Boolean expressions maintains logical consistency between coarse- and fine-grained propaganda predictions.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_dd0cc8b642d8ed92
2020.emnlp-main.320
2,020
Class representations derived from literal definitions of propaganda techniques in natural language regularize model parameters for propaganda detection.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_a83e8d90c57e13da
2020.emnlp-main.472
2,020
Micro-Dialect Identification (MDI) is a task for predicting fine-grained dialect varieties at the city level from a single short message.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_b1b218fa328d0fa7
2020.emnlp-main.472
2,020
MARBERT predicts micro-dialects at the city level from a single short message.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_f4d8b6a689e5e6d9
2020.emnlp-main.472
2,020
Spatially and linguistically-motivated multi-task learning models identify micro-dialects.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_751cf333897a8157
2020.emnlp-main.472
2,020
A large-scale dataset of low-resource Arabic micro-varieties is suited for micro-dialect identification.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_e94978b6487123c0
2020.emnlp-main.393
2,020
Leaderboards and NLP practitioners are consumers of models, and the benefit from a model is its utility to them.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_15dbe986d652c160
2020.emnlp-main.393
2,020
Leaderboards in their current form can be poor proxies for the NLP community at large.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_8ae564f69c490d9e
2020.emnlp-main.393
2,020
Leaderboards should report statistics of practical concern, such as model size, energy efficiency, and inference latency, to enable practitioners to estimate a model's utility.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_4c0c9700b362d5cf
2021.emnlp-main.691
2,021
Clause recommendation aids and accelerates contract authoring.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_694475eadf3457fc
2021.emnlp-main.691
2,021
ClauseRec uses a two-staged pipeline that predicts clause type relevance and then recommends clauses based on contract context.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_5dc85eaf7aa65c3f
2021.emnlp-main.691
2,021
BERT is pre-trained on a library of clauses with two additional tasks for clause type relevance prediction and clause recommendation.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_6173e3846afaaafe
2025.emnlp-main.603
2,025
LIDDiA navigates the drug discovery process in silico.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_726311fff70b7376
2025.emnlp-main.603
2,025
LIDDiA generates molecules meeting key pharmaceutical criteria for the majority of tested clinically relevant targets.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_917c0c787b9e94a8
2025.emnlp-main.603
2,025
LIDDiA balances exploration and exploitation in the chemical space.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_360efbc811a201b1
2025.emnlp-main.603
2,025
LIDDiA identifies a promising novel candidate molecule for the AR/NR3C4 target.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_158953a850313f8d
2025.emnlp-main.59
2,025
VC4VG optimizes video captions for text-to-video generation.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ec7d45c9195335d5
2025.emnlp-main.59
2,025
VC4VG-Bench evaluates video captions for text-to-video generation with fine-grained, multi-dimensional, and necessity-graded metrics.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_6f1fce78436a45e4
2025.emnlp-main.59
2,025
Improved caption quality correlates with better video generation performance for text-to-video generation.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_1daa9fa97a562def
2021.emnlp-main.713
2,021
Finetuned prompt-based models for sentence pair classification tasks adopt inference heuristics based on lexical overlap, such as incorrectly assuming a sentence pair has the same meaning because they consist of the same set of words.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_b9a6dffa01ff7f46
2021.emnlp-main.713
2,021
The inference heuristic based on lexical overlap is less present in zero-shot prompt-based models for sentence pair classification tasks than in finetuned prompt-based models.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_cb337177a46a8d4a
2021.emnlp-main.713
2,021
Finetuning prompt-based models for sentence pair classification tasks is destructive to useful knowledge learned during pretraining.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_2b96731a612cb34d
2021.emnlp-main.713
2,021
Adding a regularization that preserves pretraining weights mitigates the destructive tendency of few-shot finetuning in prompt-based models for sentence pair classification tasks.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_35f164cf78ebda4c
2024.emnlp-main.10
2,024
Hateful Word in Context Classification (HateWiC) determines whether a word is used hatefully in a specific context.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_4996853d82e0211c
2024.emnlp-main.10
2,024
A dataset of approximately 4000 Word-in-Context instances for classifying hateful word usage in context, each labeled by three annotators.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_57d9a256dc4851a7
2024.emnlp-main.10
2,024
Hatefulness of a word in context does not always derive from the sense definition alone.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_f9746b6d41c452bb
2024.emnlp-main.10
2,024
Including dictionary definitions improves prediction of hateful word senses, except when hateful connotations vary.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_08f207aea44cf2e8
2024.emnlp-main.10
2,024
Including annotator demographics mitigates performance drop in subjective hate prediction.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_9bebb4eb010c7bc0
2023.emnlp-main.1037
2,023
MultiCochrane is the first sentence-aligned multilingual text simplification dataset for the medical domain in English, Spanish, French, and Farsi.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_f170eefe9fc13d58
2023.emnlp-main.1037
2,023
Fine-tuned and zero-shot models can generate viable simplified medical texts in multiple languages.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_cb0c6ebb5d24d5a1
2023.emnlp-main.1037
2,023
Multilingual medical text simplification faces several outstanding challenges.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_58ccec8cff158655
2024.emnlp-main.386
2,024
Dynamic weighting by discriminator gradient magnitudes combines multiple style objectives in a reward function for multi-style controllable text generation.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_e0ef2de59b7618c0
2024.emnlp-main.386
2,024
Dynamic weighting by discriminator gradient magnitudes outperforms static weighting approaches for style control while maintaining linguistic quality.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_6c0c22d9375ac0cf
2023.emnlp-main.546
2,023
An affinity scoring function predicts transfer gain between task pairs using their task embeddings.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ce2295e271a9e864
2023.emnlp-main.546
2,023
Soft prompts serve as task embeddings that summarize task-specific information.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0a85e9a087d7432b
2023.emnlp-main.546
2,023
An affinity scoring function predicts the most transferable source tasks for a novel target task without brute-force search.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_e5de1eea1c14e4a4
2022.emnlp-main.703
2,022
DiscoSense evaluates commonsense reasoning through understanding a wide variety of discourse connectives.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_b6d140a126e375a0
2022.emnlp-main.703
2,022
Conditional Adversarial Filtering extends Adversarial Filtering by employing conditional generation to create compelling distractors.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_024d16bd55a889a6
2022.emnlp-main.703
2,022
State-of-the-art pre-trained language models perform poorly on DiscoSense.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_8250e78d1a7b9061
2022.emnlp-main.744
2,022
KOLD is a Korean offensive language dataset with hierarchical annotations of offensive language type and target, offensive span annotations, and context from article and video titles.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ff3c9760193d7948
2022.emnlp-main.744
2,022
The target group distribution in KOLD differs drastically from that in existing English offensive language datasets.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0419e45ea76ef58f
2022.emnlp-main.744
2,022
Providing context information from article and video titles improves model performance for offensiveness detection, target classification, and target group classification.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_6ec3577b0cfbf75c
2022.emnlp-main.19
2,022
MVQG is a dataset for generating engaging questions from multiple images.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_b4f074fe76cd2008
2022.emnlp-main.19
2,022
Building stories behind an image sequence enables models to generate engaging questions.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_3849c91d5240f63f
2024.emnlp-main.7
2,024
A novel framework for Avalon gameplay features a multi-agent system that facilitates efficient communication and interaction among LLM-based agents.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_7165b67fa4639bdf
2021.emnlp-main.651
2,021
Humans are capable of generating a substantial amount of adversarial examples using semantics-preserving word substitutions.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_dca64e05ae788090
2021.emnlp-main.651
2,021
Human-generated adversarial examples are not more able than the best algorithms to generate natural-reading, sentiment-preserving examples.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_caf5994597cd28ba
2021.emnlp-main.651
2,021
Human generation of adversarial examples is much more computationally efficient than the best algorithms.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_af2b75d5dc141f4e
2020.emnlp-main.189
2,020
BLANC uses context prediction as an auxiliary task in multi-task learning.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_88680137244093c8
2020.emnlp-main.189
2,020
BLANC uses a block attention method to learn the context prediction task.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_4c32b3ac2670a960
2020.emnlp-main.189
2,020
BLANC outperforms state-of-the-art QA models in reading comprehension, and this performance advantage increases as the number of answer text occurrences in the passage increases.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_c95f4c5519c5e5b3
2020.emnlp-main.189
2,020
BLANC outperforms baseline models in a zero-shot setting for supporting fact prediction on HotpotQA after training on SQuAD.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_9c7483256ff7bb28
2023.emnlp-main.501
2,023
Dialogue-guided Chain-of-Thought (DialCoT) decomposes complex reasoning problems into simpler sub-questions and generates intermediate reasoning steps in a dialogue format for Smaller Language Models.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_8eb920dadef4060d
2023.emnlp-main.501
2,023
Proximal Policy Optimization optimizes Smaller Language Models to select the optimal reasoning path.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_9ea9b18c143b9de6
2025.emnlp-main.355
2,025
DICE uses small language models to refine large language model outputs via chain-of-thought correction.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_f514cf53799d609d
2025.emnlp-main.355
2,025
DICE decouples reasoning and output specification adherence by first generating natural language responses with large language models and then refining them with small language models.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_c949a70b83ca2025
2025.emnlp-main.355
2,025
DICE constructs structured chain-of-thought adaptation datasets using a two-stage method.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_606559b07ceb4428
2025.emnlp-main.355
2,025
DICE applies a dual-tuning strategy to fine-tune small language models for generating structured outputs in an analyze-then-answer pattern.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_a752ef7d35e8c73e
2025.emnlp-main.355
2,025
DICE achieves state-of-the-art performance in format accuracy and content correctness for structured reasoning tasks.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_d57b38afbe9d76dc
2021.emnlp-main.576
2,021
Visual text representations create continuous vocabularies for machine translation by processing visually rendered text with sliding windows.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_2b5fb705b23a1898
2021.emnlp-main.576
2,021
Models using visual text representations achieve comparable performance to traditional text models on small and large machine translation datasets.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_d72774fd2e3ac477
2021.emnlp-main.576
2,021
Models using visual text representations exhibit significant robustness to varied types of noise in machine translation.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ef3d7ef6367c19ac
2021.emnlp-main.17
2,021
The partition filter network decomposes feature encoding for joint entity and relation extraction into partition and filter steps.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_f6408a92486532fd
2021.emnlp-main.17
2,021
The partition filter network uses entity and relation gates to segment neurons into two task-specific partitions and one shared partition.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_da2a386416e73940
2021.emnlp-main.17
2,021
The shared partition in the partition filter network represents inter-task information and is evenly shared between entity and relation tasks to ensure two-way interaction.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_7c58e1441280f3fa
2021.emnlp-main.17
2,021
The task-specific partitions in the partition filter network represent intra-task information and are formed through concerted efforts of both gates to make entity feature encoding and relation feature encoding dependent on each other.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_001a7a97ec7f951d
2021.emnlp-main.17
2,021
Relation prediction contributes non-negligibly to named entity prediction in joint entity and relation extraction.
4
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_799fb2d9ee4a2fa4
2025.emnlp-main.1317
2,025
Performance degradation in Multimodal Large Language Models from corrupted training data is largely reversible.
0
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_0165a962c2f78955
2025.emnlp-main.1317
2,025
Disabling a small subset of parameters in corrupted Multimodal Large Language Models almost fully restores performance.
1
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_ec8546b4bd6a5a22
2025.emnlp-main.1317
2,025
Corrupted Multimodal Large Language Models inherently differentiate between clean and corrupted samples, facilitating dataset cleaning without external intervention.
2
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00
clm_7ec4e0776cd3ea12
2025.emnlp-main.1317
2,025
A corruption-robust training paradigm mitigates the effects of corrupted data in Visual Instruction Tuning.
3
qwen/qwen3-235b-a22b-thinking-2507
f9a9b434c72c
extract_20260524T085601Z
2026-05-24T08:56:02.723585+00:00