Implicit relational attention network for few-shot knowledge graph completion

Abstract

Knowledge Graphs can not contain all the knowledge during the construction process, so needs to be completed to enhance its integrity. In real knowledge graphs, different relationships often show apparent long-tail distributions, i.e., many relationships have only a small number of entity pairs. Therefore, it is an urgent need to study few-shot knowledge graph completion. Existing methods generally complete the knowledge graph by learning representations of entities and relationships, but ignore the impact of the similarity of neighbor relations between triple entity pairs on completion. In this paper, we propose an implicit relational attention network to address this limitation. First, we propose a heterogeneous entity and relational encoder to mine one-hop neighbor information and enhance entity and relational representations through attention mechanism and convolution. Next, we propose an implicit relationship aware encoder to mine the neighbor relationship similarity information of triple entity pairs and obtain the triple dynamic relationship representation. Then we propose an adaptive relationship fusion network, which fuses the triple dynamic relationship representation and the original information of the neighbor relationship similarity of entity pairs, enhances the relationship representation of the query set to the reference set, so as to improve the accuracy of the few-shot knowledge graph completion. On two benchmark datasets, by comparing with well-known completion methods, the experimental results show that the proposed method achieves very competitive performance.