Enhancing explainability in multi-hop QA with edge addition and removal

Abstract

Multi-hop question answering is a typical task of aggregating information from multiple paragraphs to find the correct answer to a given question. Existing models use graph structures to explicitly and effectively convey information between multiple paragraphs. In addition to existing models that have improved performance using graphs, we use the intuitive characteristics of graphs to improve the explainability for multi-hop question answering. Given an original model’s graph, we add necessary edges or remove unnecessary edges to construct an explainable graph. Explainable graph has less edges and paths from the question to the answer than original graph, resulting in enhancing the explainability.