Dynamic Semantic Graph Construction and Reasoning for Explainable Multi-hop Science Question Answering

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1044–1056 August 1–6, 2021. ©2021 Association for Computational Linguistics 1044 Dynamic Semantic Graph Construction and Reasoning for Explainable Multi-hop Science Question Answering∗ Weiwen Xu, Huihui Zhang, Deng Cai and Wai Lam The Chinese University of Hong Kong {wwxu,hhzhang,wlam}@se.cuhk.edu.hk thisisjcykcd@gmail.com Abstract Knowledge retrieval and reasoning are two key stages in multi-hop question answering (QA) at web scale. Existing approaches suf- fer from low confidence when retrieving evi- dence facts to fill the knowledge gap and lack transparent reasoning process. In this paper, we propose a new framework to exploit more valid facts while obtaining explainability for multi-hop QA by dynamically constructing a semantic graph and reasoning over it. We em- ploy Abstract Meaning Representation (AMR) as semantic graph representation. Our frame- work contains three new ideas: (a) AMR-SG, an AMR-based Semantic Graph, constructed by candidate fact AMRs to uncover any hop relations among question, answer and multi- ple facts. (b) A novel path-based fact an- alytics approach exploiting AMR-SG to ex- tract active facts from a large fact pool to an- swer questions. (c) A fact-level relation mod- eling leveraging graph convolution network (GCN) to guide the reasoning process. Re- sults on two scientific multi-hop QA datasets show that we can surpass recent approaches including those using additional knowledge graphs while maintaining high explainability on OpenBookQA and achieve a new state-of- the-art result on ARC-Challenge in a computa- tionally practicable setting. 1 Introduction Multi-hop QA is one of the most challenging tasks that benefits from explainability as it mimics the human question answering setting, where multi- hop QA requires both the collection of information from large external knowledge resources and the aggregation of retrieved facts to answer complex natural language questions (Yang et al., 2018). ∗ The work described in this paper is substantially sup- ported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14204418). Question: Predators eat __. Answer Choice: bunnies Hypothesis: Predators eat bunnies. Fact 1: A bunny is a small rabbit. Fact 2: Most predators eat rabbits. eat-01 predator bunny ARG0 ARG1 Ra rabbit-01 small bunny mod domain most quant eat-01 predator Ra rabbit-01 ARG0 ARG1 Figure 1: The AMR of the hypothesis (black), Fact 1 and Fact 2. A hypothesis is a statement derived from a question and a choice. The hypothesis AMR can be inferred by relevant fact AMRs. Currently, external knowledge is mostly stored in two forms – textual and graph structure (e.g. Knowledge Graph (KG)). Textual corpora contain rich and diverse evidence facts, which are ideal knowledge resources for multi-hop QA. Especially with the success of pretrained models (Devlin et al., 2019; Liu et al., 2019; Lan et al., 2019), we can get powerful representations for such textual facts. However, retrieving relevant and useful facts to fill the knowledge gap for inferring the answer is still a challenging problem. In addition, the reasoning process over the facts is hidden by the unexplain- able neural …

Content truncated to 500 words