Dynamic knowledge graph reasoning based on deep reinforcement learning

Abstract

Knowledge graph reasoning is a task of reasoning new knowledge or conclusions based on existing knowledge. Recently, reinforcement learning has become a new technical tool for knowledge graph reasoning. However, most previous work focuses on the short fixed-step multi-hop reasoning or the single-step reasoning. In this paper, a dynamic knowledge graph reasoning framework is proposed based on deep reinforcement learning, which learns to navigate the graph to find the promising target answer conditioned on the input query. A novel reward function is constructed based on the proposed dynamic reasoning hypothesis, and the dynamic reward is proposed to obtain the dynamic reasoning model. The judgment condition is established to achieve dynamic reasoning, and an embedding model is chosen as the pre-trained model for initialization. The experimental analysis on several typical datasets shows that our model is indeed effective, and can improve over existing path-based knowledge graph reasoning models, and can be further applied to longer reasoning paths with good results.