AdaLoGN: Adaptive Logic Graph Network for Reasoning-Based Machine Reading Comprehension

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7147 - 7161 May 22-27, 2022 c⃝2022 Association for Computational Linguistics AdaLoGN: Adaptive Logic Graph Network for Reasoning-Based Machine Reading Comprehension Xiao Li and Gong Cheng and Ziheng Chen and Yawei Sun and Yuzhong Qu State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China {xiaoli.nju,chenziheng,ywsun}@smail.nju.edu.cn {gcheng,yzqu}@nju.edu.cn Abstract Recent machine reading comprehension datasets such as ReClor and LogiQA require performing logical reasoning over text. Con- ventional neural models are insufficient for logical reasoning, while symbolic reasoners cannot directly apply to text. To meet the challenge, we present a neural-symbolic approach which, to predict an answer, passes messages over a graph representing logical relations between text units. It incorporates an adaptive logic graph network (AdaLoGN) which adaptively infers logical relations to extend the graph and, essentially, realizes mutual and iterative reinforcement between neural and symbolic reasoning. We also implement a novel subgraph-to-node message passing mechanism to enhance context-option interaction for answering multiple-choice questions. Our approach shows promising results on ReClor and LogiQA. 1 Introduction Machine reading comprehension (MRC) has drawn much research attention. Early MRC datasets are not difficult for state-of-the-art neural methods. In- deed, BERT (Devlin et al., 2019) has outperformed humans on SQuAD (Rajpurkar et al., 2016). Re- cent datasets become more challenging. For ex- ample, ReClor (Yu et al., 2020) and LogiQA (Liu et al., 2020) require understanding and reasoning over logical relations described in text, where neu- ral methods showed unsatisfactory performance. For instance, consider the MRC task in Figure 1. The context consists of a set of textual propositions describing logical relations between elementary dis- course units (EDUs) (Mann and Thompson, 1988). For example, the first sentence describes an im- plication between two EDUs: “the company gets project A” implies that “product B can be put on the market on schedule”. With the help of propo- sitional calculus, humans can formalize proposi- tions and then apply inference rules in proposi- Context: If the company gets project A, product B can be put on the market on schedule. Product B is put on schedule if and only if the company’s fund can be normally turned over. If the company’s fund cannot be turned over normally, the development of product C cannot be carried out as scheduled. The fact is that the development of product C is carried out as scheduled. Question: This shows: Options: A. The company gets project A and product B is put on the market on schedule. B. The company does not get project A and product B is not put on the market on schedule. C. Product B is put on the market on schedule and the company’s fund is turned over normally. D. Product B is not put on the market on schedule, and the company’s fund turnover is extremely abnormal.  Figure 1: An example MRC task (adapted from a task in LogiQA). Logical connectives are highlighted in ital- ics. ✓marks the correct …

Content truncated to 500 words