Locate Then Ask: Interpretable Stepwise Reasoning for Multi-hop Question Answering

Proceedings of the 29th International Conference on Computational Linguistics, pages 1655–1665 October 12–17, 2022. 1655 Locate Then Ask: Interpretable Stepwise Reasoning for Multi-hop Question Answering Siyuan Wang1, Zhongyu Wei1,2∗, Zhihao Fan1, Qi Zhang3, Xuanjing Huang3 1School of Data Science, Fudan University, China 2Research Institute of Intelligent and Complex Systems, Fudan University, China 3School of Computer Science, Fudan University, China {wangsy18,zywei,fanzh18,qz,xjhuang}@fudan.edu.cn Abstract Multi-hop reasoning requires aggregating mul- tiple documents to answer a complex question. Existing methods usually decompose the multi- hop question into simpler single-hop questions to solve the problem for illustrating the explain- able reasoning process. However, they ignore grounding on the supporting facts of each rea- soning step, which tends to generate inaccu- rate decompositions. In this paper, we pro- pose an interpretable stepwise reasoning frame- work to incorporate both single-hop supporting sentence identification and single-hop question generation at each intermediate step, and uti- lize the inference of the current hop for the next until reasoning out the final result. We em- ploy a unified reader model for both intermedi- ate hop reasoning and final hop inference and adopt joint optimization for more accurate and robust multi-hop reasoning. We conduct exper- iments on two benchmark datasets HotpotQA and 2WikiMultiHopQA. The results show that our method can effectively boost performance and also yields a better interpretable reasoning process without decomposition supervision. 1 1 Introduction Recent years have witnessed an emerging trend in the task of multi-hop question answering. It requires the model to aggregate multiple pieces of documents (i.e., context) and perform multi- hop reasoning to infer the answer (Talmor and Be- rant, 2018; Khashabi et al., 2018). Several datasets have been introduced as benchmarks, such as Hot- potQA (Yang et al., 2018), 2WikiMultiHopQA (Ho et al., 2020) and WikiHop (Welbl et al., 2018), and the first two provide supporting facts supervi- sion to encourage models to further explain what supporting sentences lead to the prediction. The first generation of models for multi-hop question answering utilizes a one-step reader (Qiu ∗Corresponding author 1Codes are publicly available at https://github.com/ WangsyGit/StepwiseQA. [Question] Q: What city is the Marine Air Control Group 28 located in? [Context] P1: Marine Tactical Air Command Squadron 28 is a United States Marine Corps aviation command and control unit based at Marine Corps Air Station Cherry Point. They provide the 2nd Marine Aircraft Wings tactical headquarters and … P2: Marine Corps Air Station Cherry Point or MCAS Cherry Point is a United States Marine Corps airfield located in Havelock, North Carolina, United States, in the eastern part of the state. It ... P3-P10: … [Question Decomposition (Min et al., 2019)] Sub-Q1: Which Marine Air Control Group 28? Sub-Q2: What city is Marine Tactical Air Command Squadron 28 located in? [Stepwise Decomposition] Step1-S: Marine Tactical Air Command Squadron 28 is a … control unit based at Marine Corps Air Station Cherry Point. Step1-Q: Which is the base of Marine Air Control Group 28? Step2-S: Marine Corps Air Station Cherry Point ... United States Marine Corps airfield located in Havelock, …

Content truncated to 500 words