Modeling Hierarchical Reasoning Chains by Linking Discourse Units and Key Phrases for Reading Comprehension

Proceedings of the 29th International Conference on Computational Linguistics, pages 1467–1479 October 12–17, 2022. 1467 Modeling Hierarchical Reasoning Chains by Linking Discourse Units and Key Phrases for Reading Comprehension Jialin Chen1, Zhuosheng Zhang2,3, Hai Zhao2,3,∗ 1School of Mathematical Sciences, Shanghai Jiao Tong University 2Department of Computer Science and Engineering, Shanghai Jiao Tong University 3Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University sjtuchenjl@sjtu.edu.cn, zhangzs@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn Abstract Machine reading comprehension (MRC) poses new challenges over logical reasoning, which aims to understand the implicit logical relations entailed in the given contexts and perform in- ference over them. Due to the complexity of logic, logical relations exist at different granu- larity levels. However, most existing methods of logical reasoning individually focus on ei- ther entity-aware or discourse-based informa- tion but ignore the hierarchical relations that may even have mutual effects. In this paper, we propose a holistic graph network (HGN) which deals with context at both discourse level and word level, as the basis for logical reasoning, to provide a more fine-grained relation extrac- tion. Specifically, node-level and type-level relations, which can be interpreted as bridges in the reasoning process, are modeled by a hi- erarchical interaction mechanism to improve the interpretation of MRC systems. Experimen- tal results on logical reasoning QA datasets (ReClor and LogiQA) and natural language in- ference datasets (SNLI and ANLI) show the effectiveness and generalization of our method, and in-depth analysis verifies its capability to understand complex logical relations. 1 Introduction Machine reading comprehension (MRC) is a chal- lenging task that requires machines to answer a question according to given passages (Hermann et al., 2015; Rajpurkar et al., 2016, 2018; Lai et al., 2017). A variety of datasets have been in- troduced to push the development of MRC to a more complex and more comprehensive pattern, such as conversational MRC (Reddy et al., 2019; Choi et al., 2018), multi-hop MRC (Yang et al., 2018), and commonsense reasoning (Davis and Marcus, 2015; Bhagavatula et al., 2020; Talmor et al., 2019; Huang et al., 2019). In particular, ∗Corresponding author. This work was supported in part by the Key Projects of National Natural Science Foundation of China under Grants U1836222 and 61733011. Example (taken from ReClor dataset) Context: Most lecturers who are effective teachers are eccentric, but some non-eccentric lecturers are very effective teachers. In addition, every effective teacher is a good communicator. Question: Which one of the following statements follows logically from the statements above? Options: A: Most lecturers who are good communicators are eccentric. B: Some non-eccentric lecturers are effective teachers but are not good communicators. C: All good communicators are effective teachers. D: Some good communicators are eccentric. ✓ Figure 1: An example from Reclor dataset. The example mainly talks about "effective teachers, non-eccentric, eccentric, good communicator". some recent multi-choice MRC datasets pose even greater challenges to the logical reasoning abil- ity of models (Yu et al., 2020; Liu et al., 2020a) which are not easy for humans to do well, …

Content truncated to 500 words