Ask to Understand: Question Generation for Multi-hop Question Answering

Ask to Understand: Question Generation for Multi-hop Question Answering Jiawei Li, Mucheng Ren, Yang Gao �, Yizhe Yang School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China Beijing Engineering Research Center of High Volume Language Information Processing and Cloud Computing Applications, Beijing, China {jwli,renm,gyang,yizheyang}@bit.edu.cn Abstract Multi-hop Question Answering (QA) requires the machine to answer complex questions by �nd- ing scattering clues and reasoning from multiple documents. Graph Network (GN) and Ques- tion Decomposition (QD) are two common approaches at present. The former uses the �black- box� reasoning process to capture the potential relationship between entities and sentences, thus achieving good performance. At the same time, the latter provides a clear reasoning logical route by decomposing multi-hop questions into simple single-hop sub-questions. In this paper, we propose a novel method to complete multi-hop QA from the perspective of Question Genera- tion (QG). Speci�cally, we carefully design an end-to-end QG module on the basis of a classical QA module, which could help the model understand the context by asking inherently logical sub-questions, thus inheriting interpretability from the QD-based method and showing superior performance. Experiments on the HotpotQA dataset demonstrate that the effectiveness of our proposed QG module, human evaluation further clari�es its interpretability quantitatively, and thorough analysis shows that the QG module could generate better sub-questions than QD meth- ods in terms of �uency, consistency, and diversity. 1 Intorduction Unlike single-hop QA (Rajpurkar et al., 2016; Trischler et al., 2017; Lai et al., 2017) where the answers could usually be derived from a single paragraph or sentence, multi-hop QA (Welbl et al., 2018; Yang et al., 2018) is a challenging task that requires soliciting hidden information from scattered documents on different granularity levels and reasoning over it in an explainable way. The HotpotQA (Yang et al., 2018) was published to leverage the research attentions on reasoning processing and explainable predictions. Figure 1 shows an example from HotpotQA, where the question requires �rst �nding the name of the company ( Tata Consultancy Services), and then the address of the company (Mumbai). While, a popular stream of Graph Network-based (GN) approaches (De Cao et al., 2019; Tu et al., 2019; Ding et al., 2019; Fang et al., 2020) was proposed due to the structures of scattered evidence could be captured by the graphs and re�ected in the representing vectors. However, the reasoning process of the GN-based method is entirely different from human thoughts. Speci�cally, GN tries to �gure out the underlying relations between the key entities or sentences from the context. However, the process is a �black-box�; we do not know which nodes in the network are involved in reasoning for the �nal answer, thus showing relatively poor interpretability. Inspired by that human solves such questions by following a transparent and explainable logical route, another popular stream of Question Decomposition-based (QD) approaches became favored in recent years (Fu et al., 2021; Nishida et al., 2019; Min et al., 2019; Jiang and Bansal, 2019b). The method mim- ics human reasoning to decompose …

Content truncated to 500 words