Multi-hop question answering for SRLGRN augmented by textual relationship modelling

Abstract

Multi-hop question answering aims to predict answers to questions and generate supporting facts for answers by reasoning over the content of multiple documents. The recently proposed Semantic Role Labeling Graph Reasoning Network (SRLGRN) has achieved excellent performance on multi-hop QA tasks. However, SRLGRN is lacking in modelling the textual relationships, which are import cues for reasoning. To this end, this paper proposes an enhanced SRLGRN multi-hop question answering approach by modelling textual relationships at different granularity (document relationships and sentence relationships). By modelling document relationships, a novel document filter based on document relationship threshold is designed for SRLGRN to dynamically select documents relevant to the question from multiple documents. By modelling sentence relationships, a sentence relationship-aware answer type prediction module is added to SRLGRN, which models sentences in documents as sentence graphs and then uses graph convolution network to predict answer type. The obtained answer type further guide the answer reasoning module of SRLGRN to obtain question answer with supporting facts. The experimental results show that the proposed scheme outperforms SRLGRN in terms of answer prediction and supporting fact prediction, with a 2% improvement in answer F1 metrics and a 3.1% improvement in joint F1 performance.