Speaker: Zhenyun Deng, PhD student, supervised by Michael Wittbrock, and Pat Riddle
Abstract: Multi-hop question has become an important challenge in reading comprehension (RC) because it requires integrating information from scattered texts across multiple paragraphs. In this paper, we propose a Knowledge and Information Fusion Graph Mask Network (KIF-GMN) to perform reasoning for multi-hop questions. To avoid missing information related to the question, we propose to integrate the information of different levels into a graph and construct edges for nodes that have a semantic relationship by Open Information Extraction (Open IE) and Coreference Resolution (CR). Moreover, adding a large number of nodes and edges to the graph will lead to inefficient information integration in the process of graph reasoning and cause scalability issues to the reasoning chain finding. The KIF-GMN also designs dynamic mask modules for the graph to filter out nodes that are irrelevant to the question.