Path Reasoning over Knowledge Graph: A Multi-agent and Reinforcement Learning Based Method

Abstract

Relation reasoning over knowledge graphs is an important research problem in the fields of knowledge engineering and artificial intelligence, because of its extensive applications (e.g., knowledge graph completion and question answering). Recently, reinforcement learning has been successfully applied to multi-hop relation reasoning (i.e., path reasoning). And, a kind of practical path reasoning, in the form of query answering (e.g., (entity, relation, ?)), has been proposed and attracted much attention. However, existing methods for such type of path reasoning focus on relation selection and underestimate the importance of entity selection during the reasoning process. To solve this problem, we propose a Multi-Agent and Reinforcement Learning based method for Path Reasoning, thus called MARLPaR, where two agents are employed to carry out relation selection and entity selection, respectively, in an iterative manner, so as to implement complex path reasoning. Experimental comparison with the state-of-the-art baselines on two benchmark datasets validates the effectiveness and merits of the proposed method.