KLUE constructs 8 Korean NLU tasks, using from-scratch data collection and pretrained models KLUE-BERT and KLUE-RoBERTa, outperforming multilingual and open-source baselines.

Sungjoon Park, Jihyung Moon, Sungdong Kim et al.

TAT-QA dataset combines tables and text; TAGOP model improves F1 score to 58.0%.

Fengbin Zhu, Wenqiang Lei, Youcheng Huang et al.

Kleister datasets challenge existing KIE models; best model achieved 81.77% and 83.57% F1 on NDA and Charity datasets.

Tomasz Stanisławek, Filip Graliński, Anna Wróblewska et al.

Survey on deep learning dialogue systems, covering model types, system types, evaluation methods, and emerging trends.

Jinjie Ni, Tom Young, Vlad Pandelea et al.

Inter-GPS integrates formal language and symbolic reasoning, constructing Geometry3K, significantly improving geometric problem solving accuracy.

Pan Lu, Ran Gong, Shibiao Jiang et al.

D2S system integrates query-based long-form QA and dense vector retrieval for automatic scientific paper-to-slide generation.

Edward Sun, Yufang Hou, Dakuo Wang et al.

DExperts combines expert and anti-expert models to control text generation at decoding time, significantly reducing GPT-3 toxicity.

Alisa Liu, Maarten Sap, Ximing Lu et al.

Scaling XLM-R to 3.5B and 10.7B parameters significantly improves cross-lingual understanding, surpassing previous models on benchmarks.

Naman Goyal, Jingfei Du, Myle Ott et al.

Introduces FRANK benchmark with fine-grained factuality error classification, evaluating summarization models and metrics.

Artidoro Pagnoni, Vidhisha Balachandran, Yulia Tsvetkov

Proposes EW and MW metrics to separately evaluate language models' syntactic systematicity and behavior, revealing overestimation of their capabilities.

Benjamin Newman, Kai-Siang Ang, Julia Gong et al.

Detect out-of-distribution data for pretrained Transformers using contrastive loss and Mahalanobis distance, achieving near-perfect performance.

Wenxuan Zhou, Fangyu Liu, Muhao Chen

Introduces the concept of knowledge neurons using integrated gradients, identifying neurons responsible for factual knowledge in BERT, enabling knowledge editing without fine-tuning, with an average of 4.13 neurons per fact.

Damai Dai, Li Dong, Yaru Hao et al.

Introduced Domain Conditional Pointwise Mutual Information to address surface form competition, enhancing zero-shot task accuracy.

Ari Holtzman, Peter West, Vered Shwartz et al.

Proposes Condenser, a pre-training architecture with conditional dense representations, significantly improving retrieval tasks.

Luyu Gao, Jamie Callan

Retrieval-augmented architecture reduces hallucination by over 60%, outperforming SOTA on two knowledge-grounded tasks, enhancing factual accuracy in dialogue.

Kurt Shuster, Spencer Poff, Moya Chen et al.

Integrating source and target context modeling, including coreference resolution and multi-hop reasoning via Transformer-XH, improves detection of previously fact-checked claims by over 10 MAP points.

Shaden Shaar, Firoj Alam, Giovanni Da San Martino et al.

SummScreen dataset for abstractive screenplay summarization, challenging due to dialogue and plot details.

Mingda Chen, Zewei Chu, Sam Wiseman et al.

AR-LSAT combines Transformer and symbolic reasoning, revealing deep reasoning gaps in current models with accuracy below 35%. ARM achieves 65%, still behind humans.

Wanjun Zhong, Siyuan Wang, Duyu Tang et al.

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.