The 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020) has announced the list of accepted papers. As appeared on Synced review first, this year’s ACL edition accepted a total of 779 papers out of 3,088 submissions. These papers are based on substantial, original, and unpublished research in all aspects of Computational Linguistics and Natural Language Processing.

Here a few exciting works that were accepted for ACL 2020: –

Improving Multi-Hop Question Answering (QA) From IISc Bangalore

In this paper, the authors from IISc Bangalore, proposed EmbedKGQA to effectively perform multi-hop KGQA over sparse knowledge graphs (KGs). This method also overcomes the requirement of answer selection from a prespecified neighbourhood, which is a constraint enforced by previous multi-hop KGQA methods. The results from the extensive experiments on multiple benchmark datasets, show that EmbedKGQA is more effective than other state-of-the-art baselines.

NILE From IISc Bangalore

In this work, the authors focus on the task of natural language inference (NLI) and investigate if NLI systems which produce labels with high accuracy could be built while generating faithful explanations of its decisions? Based on this notion, they propose NILE — Natural Language Inference Over Label-Specific Explanations, a novel NLI method that utilises auto-generated label-specific NL explanations to produce labels along with its faithful explanation.


Language models are typically trained on free-form natural language text, hence may not be suitable for tasks like semantic parsing over structured data, which require reasoning over both free-form NL questions and structured tabular data (e.g., database tables). In this paper, the authors from Carnegie Mellon University and Facebook AI Research present TABERT, a pretrained LM that jointly learns representations for NL sentences and (semi-) structured tables. TABERT is trained on a large corpus of 26 million tables and their English contexts.

Emotion Helps Sentiment From IIT Patna

In this paper, the authors from IIT Patna and Kyoto University, Japan propose a two-layered multi-task attention-based neural network that performs sentiment analysis through emotion analysis. The proposed approach is based on Bidirectional Long Short-Term Memory (Bi-LSTM) and uses Distributional Thesaurus as a source of external knowledge to improve sentiment and emotion prediction. 

Optimising The Factual Correctness Of A Summary From Stanford University

In this work, the authors develop a general framework to evaluate the factual correctness of a generated summary by fact-checking it automatically against its reference using an information extraction module. They also propose a training strategy that optimises a neural summarisation model with a factual correctness reward via reinforcement learning. With the experiments on two separate datasets collected from hospitals, show both automatic and human evaluation that the proposed approach substantially improves the factual correctness and overall quality of outputs.

Null It Out From Bar Ilan University & Allen Institute Of AI 

In this paper, the authors present Iterative Null-Space Projection (INLP), a novel method based on repeated training of linear classifiers that predicts a certain property. Using this method, the authors try to remove information from representations by projecting them on their null-space. They show that this method is able to mitigate bias in word embeddings, as well as to increase fairness in a setting of multi-class classification.

Unsupervised Domain Clusters In Pretrained Language Models By Roee Aharoni & Yoav Goldberg

The authors show that massive pre-trained language models implicitly learn sentence representations that cluster by domains without supervision, suggesting a simple data-driven definition of domains in textual data. They propose domain data selection methods that require only a small set of in-domain monolingual data. The authors write that their method outperforms state-of-the-art methods.

Learning To Deceive With Attention-Based Explanations From CMU

In this work, the authors from Carnegie Mellon University and Twitter, introduce a method that diminishes the total weight assigned to designated impermissible tokens, even when the models can be shown to rely on these features to drive predictions. Across multiple models and tasks, this approach manipulates attention weights while staying accurate. The results cast doubt on attention’s reliability as a tool for auditing algorithms in the context of fairness and accountability.

Predicting Performance For Natural Language Processing Tasks By CMU

In this work, the authors attempt to explore how well an NLP model can perform under an experimental setting, without actually training or testing the model. To do so, they built regression models to predict the evaluation score of an NLP experiment given the experimental settings as input. Experimenting on 9 different NLP tasks, they find that the predictors can produce meaningful predictions over unseen languages and different modeling architectures, outperforming reasonable baselines as well as human experts.

Staying True To Your Word By Martin Tutek And Jan Snajder

In this paper, the authors back up the criticisms around ‘attention’ in language models. They provided an explanation as to why attention has seen rightful critique when used with recurrent networks in sequence classification tasks. They also propose a remedy to these issues in the form of a word-level objective, and the findings give credibility for attention to provide faithful interpretations of recurrent models.

Spelling Error Correction From ByteDance AI Lab

In this paper, the authors introduce a state-of-the-art method for the task that selects a character from a list of candidates for correction (including non-correction) at each position of the sentence on the basis of BERT. They propose a novel neural architecture to address the problem, which consists of a network for error detection and a network for error correction based on BERT. Experimental results show that the performance of the proposed method is significantly better than the baselines, including the one solely based on BERT.

Simultaneous Translation Policies From Baidu, USA

In this work, the authors from Baidu, USA and Oregon State University, introduce an algorithm that achieves adaptive policies via a simple heuristic composition of a set of fixed policies. Their experiments on Chinese→English and German→English show that adaptive policies can outperform fixed ones for the same latency, and even surpasses the state-of-the-art full-sentence translation methods with much lower latency.

Check out the full list of accepted papers here.