Iterative Reasoning over Knowledge Graph

The concept reasoning is an essential task in text data management and understanding. Recent methods usually capture shallow semantic features and cannot extend to multi-hop reasoning. Knowledge graphs have rich text information and connections. We use a knowledge graph to encode complex semantic relation between evidence and question. The nodes represent valuable information as clue entities and candidate answers in evidence and question, and the edges represent the reasoning rules between nodes. In this paper, we propose a graph-based reasoning framework with iterative steps. The model obtains the completed evidence chain through iterative reasoning. The new approach iteratively infers the clue entities and candidate answers from the question and clue paragraphs to as new nodes to expand the semantic relation graph. Then we update the semantic representation of the questions and context via memory network and apply the graph attention network to encode the reasoning paths in the knowledge graph. Extensive experiments on commonsense reasoning and multi-hop question answering verified the advantage and improvements of the proposed approach.