Explore, Propose, and Assemble: An Interpretable Model for Multi-Hop Reading Comprehension

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2714–2725 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2714 Explore, Propose, and Assemble: An Interpretable Model for Multi-Hop Reading Comprehension Yichen Jiang∗ Nitish Joshi∗ Yen-Chun Chen Mohit Bansal UNC Chapel Hill {yichenj, nitish, yenchun, mbansal}@cs.unc.edu Abstract Multi-hop reading comprehension requires the model to explore and connect relevant infor- mation from multiple sentences/documents in order to answer the question about the con- text. To achieve this, we propose an in- terpretable 3-module system called Explore- Propose-Assemble reader (EPAr). First, the Document Explorer iteratively selects relevant documents and represents divergent reasoning chains in a tree structure so as to allow assim- ilating information from all chains. The An- swer Proposer then proposes an answer from every root-to-leaf path in the reasoning tree. Finally, the Evidence Assembler extracts a key sentence containing the proposed answer from every path and combines them to pre- dict the final answer. Intuitively, EPAr ap- proximates the coarse-to-fine-grained compre- hension behavior of human readers when fac- ing multiple long documents. We jointly op- timize our 3 modules by minimizing the sum of losses from each stage conditioned on the previous stage’s output. On two multi-hop reading comprehension datasets WikiHop and MedHop, our EPAr model achieves significant improvements over the baseline and compet- itive results compared to the state-of-the-art model. We also present multiple reasoning- chain-recovery tests and ablation studies to demonstrate our system’s ability to perform in- terpretable and accurate reasoning.1 1 Introduction The task of machine reading comprehension and question answering (MRC-QA) requires the model to answer a natural language question by finding relevant information and knowledge in a given natural language context. Most MRC ∗equal contribution; part of this work was done during the second author’s internship at UNC (from IIT Bombay). 1Our code is publicly available at: https://github.com/jiangycTarheel/EPAr datasets require single-hop reasoning only, which means that the evidence necessary to answer the question is concentrated in a single sentence or lo- cated closely in a single paragraph. Such datasets emphasize the role of locating, matching, and aligning information between the question and the context. However, some recent multi-document, multi-hop reading comprehension datasets, such as WikiHop and MedHop (Welbl et al., 2017), have been proposed to further assess MRC sys- tems’ ability to perform multi-hop reasoning, where the required evidence is scattered in a set of supporting documents. These multi-hop tasks are much more chal- lenging than previous single-hop MRC tasks (Ra- jpurkar et al., 2016, 2018; Hermann et al., 2015; Nguyen et al., 2016; Yang et al., 2015) for three primary reasons. First, the given context contains a large number of documents (e.g., 14 on aver- age, 64 maximum for WikiHop). Most existing QA models cannot scale to the context of such length, and it is challenging to retrieve a reason- ing chain of documents with complete informa- tion required to connect the question to the an- swer in a logical way. Second, given a reason- …

Content truncated to 500 words