Multi-choice Relational Reasoning for Machine Reading Comprehension

Proceedings of the 28th International Conference on Computational Linguistics, pages 6448–6458 Barcelona, Spain (Online), December 8-13, 2020 6448 Multi-choice Relational Reasoning for Machine Reading Comprehension Wuya Chen1, Chunyu Kit2, Xiaojun Quan1∗, Zhengcheng Min1, Jiahai Wang1 1School of Data and Computer Science, Sun Yat-sen University, Guangzhou 2Department of Linguistics and Translation, City University of Hong Kong, Hong Kong quanxj3@mail.sysu.edu.cn, ctckit@cityu.edu.hk Abstract This paper presents our study of cloze-style reading comprehension by imitating human reading comprehension, which normally involves tactical comparing and reasoning over candidates while choosing the best answer. We propose a multi-choice relational reasoning (McR2) model with an aim to enable relational reasoning on candidates based on fusion representations of document, query and candidates. For the fusion representations, we develop an efficient encoding architec- ture by integrating the schemes of bidirectional attention flow, self-attention and document-gated query reading. Then, comparing and inferring over candidates are executed by a novel relational reasoning network. We conduct extensive experiments on four datasets derived from two public corpora, Children’s Book Test and Who DiD What, to verify the validity and advantages of our model. The results show that it outperforms all baseline models significantly on the four bench- mark datasets. The effectiveness of its key components is also validated by an ablation study. 1 Introduction Machine reading comprehension (MRC) is a challenging task that requires much semantic understanding and reasoning using various clues from texts (Seo et al., 2016). Its general form is to ask a computer to answer questions in natural language according to its understanding of a given article or a context. As a specific form of MRC, cloze-style reading comprehension has recently gained increasing attention. Cloze-style MRC is a task to fill in a blank in the query with an appropriate word or phrase according to given context.Several large-scale datasets (Hill et al., 2016; Hermann et al., 2015; Onishi et al., 2016) for this task have been released, facilitating the development of various machine learning models (Kadlec et al., 2016; Trischler et al., 2016; Dhingra et al., 2017; Ghaeini et al., 2018). Most of these learning systems are built upon multi-hop architectures and attention mechanisms (Dhingra et al., 2017), which have been shown to excel at distilling useful information and learning the “importance” distribution over inputs. There are also some works that mimic the cognitive process of human reasoning with complicated hypothesis testing frameworks (Trischler et al., 2016; Munkhdalai and Yu, 2016). However, none of these models provide explicit reasoning among candidate answers with respect to a given context and query.As a practical skill for cloze test, humans often need to tactically compare two candidate answers while making decision, especially when more than one candidate appears to be competent. In this paper, we propose a multi-choice relational reasoning (McR2) model to imitate the above process. It first learns representations of document and query via a hierarchical multi-stage encoding architecture to explore the relations between document and query. The encoding architecture integrates the mechanisms of bidirectional attention flow (Seo et al., 2016), …

Content truncated to 500 words