QA-GNN combines relevance scoring and joint graph reasoning, outperforming LM+KG models with 4.7% improvement on CommonsenseQA.

Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut et al.

Introduces MultiModalQA dataset with 29,918 questions requiring joint reasoning over text, tables, and images; ImplicitDecomp outperforms baselines with 51.7 F1.

Alon Talmor, Ori Yoran, Amnon Catav et al.

Query-based multi-domain meeting summarization with locate-then-summarize framework; QMSum dataset highlights challenges in long meeting understanding.

Ming Zhong, Da Yin, Tao Yu et al.

Introduced StylePTB, a large-scale benchmark with 21 fine-grained style changes, advancing controllable text style transfer research.

Yiwei Lyu, Paul Pu Liang, Hai Pham et al.

Meta-tuning on combined datasets boosts large model zero-shot classification by 6.3% (220M to 770M params), surpassing prior SOTA.

Ruiqi Zhong, Kristy Lee, Zheng Zhang et al.

Dynabench introduces a dynamic adversarial data collection platform, significantly reducing model error rates across four NLP tasks by leveraging multi-round human-model interactions.

Douwe Kiela, Max Bartolo, Yixin Nie et al.

HEPOS employs head-wise positional strides, reducing complexity and enabling processing of ten times longer documents for improved ROUGE scores.

Luyang Huang, Shuyang Cao, Nikolaus Parulian et al.

FeTaQA introduces a 10K Wikipedia-based dataset for complex, free-form table question answering, emphasizing multi-fact reasoning.

Linyong Nan, Chiachun Hsieh, Ziming Mao et al.

L’AMBRE combines dependency parsing and automatic rule extraction to evaluate grammatical well-formedness in multilingual generated texts, validated on MT tasks.

Adithya Pratapa, Antonios Anastasopoulos, Shruti Rijhwani et al.

Proposes whitening transformation to improve BERT sentence embeddings, boosting semantic similarity performance by 5-10 points and reducing storage by 70%.

Jianlin Su, Jiarun Cao, Weijie Liu et al.

Proposes an uncertainty-aware beam search algorithm to reduce hallucinations in conditional language generation.

Yijun Xiao, William Yang Wang

GLM introduces autoregressive blank infilling with 2D positional encoding, outperforming BERT, T5, and GPT on multiple NLP tasks.

Zhengxiao Du, Yujie Qian, Xiao Liu et al.

This study reveals that current NLP models for math word problems rely heavily on shallow heuristics, with the SVAMP challenge set exposing their fragility, dropping accuracy significantly.

Arkil Patel, Satwik Bhattamishra, Navin Goyal

CANINE is a tokenization-free character encoder using downsampling and deep transformers, improving multilingual task performance with fewer parameters.

Jonathan H. Clark, Dan Garrette, Iulia Turc et al.

Identifies fundamental challenges in LFQA evaluation and datasets; uses sparse Transformer and contrastive retrieval, achieving SOTA but with answers not truly grounded.

Kalpesh Krishna, Aurko Roy, Mohit Iyyer

Proposes RFA, a linear-time softmax approximation using random features, boosting efficiency in long sequence tasks.

Hao Peng, Nikolaos Pappas, Dani Yogatama et al.

Proposes training GPT-2 on UCI chess notation with RAP to enhance state tracking, achieving over 97% accuracy in move and piece position prediction.

Shubham Toshniwal, Sam Wiseman, Karen Livescu et al.

Probing classifiers analyze neural language models by predicting linguistic properties from internal representations, highlighting their strengths and limitations.

Yonatan Belinkov

TILT model integrates text, image, and layout for improved document understanding accuracy.

Rafał Powalski, Łukasz Borchmann, Dawid Jurkiewicz et al.

MAUVE introduces a divergence frontier-based metric to quantify the gap between neural and human text distributions, leveraging embedding quantization and area under the divergence curve for robust evaluation.

Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers et al.