RE-GCN: Relation Enhanced Graph Convolutional Network for Entity Alignment in Heterogeneous Knowledge Graphs

Abstract

Entity alignment is a fundamental task of matching synonymous entities from different knowledge graphs (KGs). Most of the existing methods perform this task by evaluating the similarity among entity embeddings learned from heterogeneous KGs, where Graph Convolutional Network (GCN) based embedding is widely adopted for capturing complex network structure. However, the semantics and directional information of relations are ignored in previous GCN based efforts, which affect the integrality of embedding definitely and decrease the efficiency consequently. To overcome this shortcoming, this paper proposes a Relation-Enhanced Graph Convolutional Network (RE-GCN) method for entity alignment including two stages. First, to take advantage of the semantics of the relations, a novel triadic graph is designed to integrate relation nodes into the primal graph by using triadic closure. In a triadic graph, both relations and entities nodes could be organized in a unified network. The corresponding triadic graph convolution is utilized together with the primal one to learn the relation and entity embeddings, simultaneously. Second, in order to make use of direction information of the relations, a bidirectional context aggregation mechanism is proposed to aggregate the embeddings from the first stage. The final aggregation embeddings are utilized for entity alignment. On three real-world multilingual datasets, experimental results demonstrate that RE-GCN produces a more excellent performance compared with some state-of-the-art entity alignment methods.