Exploiting Explicit Paths for Multi-hop Reading Comprehension

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2737–2747 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2737 Exploiting Explicit Paths for Multi-hop Reading Comprehension Souvik Kundu†∗and Tushar Khot‡ and Ashish Sabharwal‡ and Peter Clark‡ †Department of Computer Science, National University of Singapore ‡Allen Institute for Artificial Intelligence, Seattle, WA, U.S.A. souvik@comp.nus.edu.sg, {tushark,ashishs,peterc}@allenai.org Abstract We propose a novel, path-based reasoning approach for the multi-hop reading compre- hension task where a system needs to com- bine facts from multiple passages to answer a question. Although inspired by multi-hop reasoning over knowledge graphs, our pro- posed approach operates directly over unstruc- tured text. It generates potential paths through passages and scores them without any di- rect path supervision. The proposed model, named PathNet, attempts to extract implicit relations from text through entity pair repre- sentations, and compose them to encode each path. To capture additional context, Path- Net also composes the passage representations along each path to compute a passage-based representation. Unlike previous approaches, our model is then able to explain its reason- ing via these explicit paths through the pas- sages. We show that our approach outper- forms prior models on the multi-hop Wikihop dataset, and also can be generalized to apply to the OpenBookQA dataset, matching state- of-the-art performance. 1 Introduction Many reading comprehension (RC) datasets (Ra- jpurkar et al., 2016; Trischler et al., 2017; Joshi et al., 2017) have been proposed recently to eval- uate a system’s ability to answer a question from a given text passage. However, most of the ques- tions in these datasets can be answered by using only a single sentence or passage. As a result, systems designed for these tasks may not be able to compose knowledge from multiple sentences or passages, a key aspect of natural language under- standing. To remedy this, new datasets (Weston et al., 2015; Welbl et al., 2018; Khashabi et al., 2018a; Mihaylov et al., 2018) have been proposed, ∗Work performed while doing an internship at the Allen Institute for Artificial Intelligence. Query: (always breaking my heart, record label, ?) Supporting Passages: (p1) “Always Breaking My Heart” is the second sin- gle from Belinda Carlisle’s A Woman and a Man al- bum , released in 1996 ( see 1996 in music ) . It ... (p2) A Woman and a Man is the sixth studio al- bum by American singer Belinda Carlisle, released in the United Kingdom on September 23, 1996 by Chrysalis Records (then part of the EMI Group, ... Candidates: chrysalis records, emi group, virgin records, ... Answer: chrysalis records Paths: (“Always Breaking My Heart” ... single from ... A Woman and a Man) (A Woman and a Man ... released ... by ... Chrysalis Records) Figure 1: Example illustrating our proposed path ex- traction and reasoning approach. requiring a system to combine information from multiple sentences in order to arrive at the answer, referred to as multi-hop reasoning. Multi-hop reasoning has been studied for ques- tion …

Content truncated to 500 words