Dynamic Multi-hop Reasoning

Multi-hop reasoning is an essential part of the current reading comprehension and question answering areas. The reasoning methods have been extensively studied, and most of them are generally focused on the pre-retrieval based inference, with the help of a few paragraphs. These methods are fixed and unable to cope with dynamic and complex questions. Here, we propose to utilize the dynamic graph reasoning network for multi-hop reading comprehension question answering. Specifically, the new approach continuously infers the clue entities and candidate answers based on the question and clue paragraphs. The clue entities and candidate answers extracted at each hop are used as new nodes to expand the dynamic graph. Then we iteratively update the semantic representation of the questions via dynamic question memory, and apply the graph attention network to encode the information of inference paths. Extensive experiments on two datasets verify the advantage and improvements of the proposed approach.