Knowledge Graph Completion Based on Multi-Relation Graph Attention Network

Abstract

Knowledge Graph Completion (KGC), can be performed mainly by inferring missing facts from entities and relations already in the knowledge graphs. However, most methods for KGC only focus on modeling undirected or single relational graph data, ignoring semantic information of multiple relations. We argue that information in the directed edge flows bi-directionally and there exists a latent reverse relation for each relation. These latent reverse relations contain additional information for completing knowledge graphs. Thus, in this paper, we propose a novel method called Knowledge Graph Completion based on Multi-relation Graph Attention neTwork (KGC-MGAT) for more accurate knowledge graph completion. We propose a multi-relational graph attention network to embed each triple and its reverse one respectively. Finally, we design a novel optimal objective for training. We conduct extensive experiments, and the results show the superiority of our method.