Entities and Relations Aware Graph Convolutional Network for Knowledge Base Completion

Abstract

Graph Neural Networks (GNNs) have recently been shown to be quite effective in modeling graph-structured data. Recent methods such as RGCN and SACN, have achieved the most advanced results in knowledge graph completion. However, previous efforts are mostly restricted to aggregating information given by neighboring nodes only, ignoring the information given by neighboring edges. This paper proposes a novel Entities and Relations Aware Graph Convolutional Network (ERA-GCN), with an encoder-decoder framework which jointly embeds both entities and relations in a multi-relation graph. In the encoder end, ERA-GCN uses a weighted graph convolutional network to capture both graph structure and neighborhood information. In the decoder end, we utilize Conv-TransE to retain the translational property between entity and relation embedding, leading to better link prediction performance. We evaluate our proposed method on standard FB15k-237 and WNISRR datasets, and achieve about 11% relative improvement compared to current state-of-the-art ConvE in terms of HITS@l, HITS@3 and HITS@10.