PathReasoner: Explainable reasoning paths for commonsense question answering

Abstract

Commonsense question answering has attracted increasing attention as a challenging task requiring the human reasoning process of answering questions with the help of abundant commonsense knowledge. Existing methods mostly resort to large pre-trained language models and face many difficulties when dealing with the out-of-scope reasoning target, and are unaware of explainable structured information. In this paper, we explore explicitly incorporate external reasoning paths with structured information to explain and facilitate commonsense QA. For this purpose, we propose a PathReasoner to both extract and learn from such structured information. The proposed PathReasoner consists of two main components, a path finder and a hierarchical path learner. To answer a commonsense question, the path finder first retrieves explainable reasoning paths from a large-scale knowledge graph, then the path learner encodes the paths with hierarchical encoders and uses the path features to predict the answers. The experiments on two typical commonsense QA datasets demonstrate the effectiveness of the PathReasoner. The case study gives insightful findings that the reasoning paths provide explainable information for the question answering through the PathReasoner.