Entity Relation Interactive Graph Convolutional Network for Knowledge Embedding

Yang He, Bei Hui, Quanjiang Guo, Ling Tian, Dong Liu

Abstract: There are several research grown up in recent years, which focused on GCN methods for complex relational graphs. The knowledge graph with rich semantic information is a typical complex relational graph and has become a major research in the field. Many existing GCN methods for complex relational graphs try to combine knowledge embedding methods to capture the rich semantic, but they are limited to simply propagating and updating the embeddings of entities and relations, ignoring the interactions between entity and relation information. Therefore, this paper proposes a novel framework for knowledge embedding, named as entity relation interactive graph convolutional network (ER-GCN), which combines graph convolution operations and knowledge embedding methods. Specifically, ER-GCN incorporates the multi-hop neighbor nodes by constructing auxiliary edges for entity embedding, and introduces the relation information semantically similar to the given relation for relation embedding. Experimental results on three datasets show the well performance of ER-GCN on link prediction and node classification.