Learning to Answer Complex Questions with Evidence Graph

Abstract

Text-based end-to-end question answering (QA) systems have attracted more attention for their good robustness and excellent performance in dealing with complex questions. However, this kind of method lacks certain interpretability, which is essential for the QA system. For instance, the interpretability of answers is particularly significant in the medical field, in that interpretable answers are more credible and apt to acception. The methods based on knowledge graph (KG) can improve the interpretability, but suffer from the problems of incompleteness and sparseness of KG. In this paper, we propose a novel method (EGQA) to solve complex question answering via combining text and KG. We use Wikipedia as a text source to extract documents related to the question and extract triples from the documents to construct a raw graph (i.e., a small-scale KG). Then, we extract the evidence graphs from the raw graph and adopt Attention-based Graph Neural Network (AGNN) to embed them to find the answer. Our experiments conduct on a real medical dataset Head-QA, which shows that our approach can effectively improve the interpretability and performance of complex question answering.