Knowledge Graph Embedding based on Line Graph Attention Mechanism

Abstract The broad application prospects of knowledge graphs have spawned many research work in the field of knowledge graph completion. In these works, models based on convolution neural networks perform well on link prediction tasks by generating expressive feature embedding. Some models based on graph neural networks also obtain good knowledge graph embedding results by capturing graph structure features and node features in the knowledge graph. However, these models often rely more on the entity features in the triples and ignore the importance of the relationship features, so they cannot guarantee the generalization ability on the triples containing unseen entities. To this end, we propose a method of learning Knowledge Graph embedding on Line Graphs using Graph Attention mechanism (KG-LGAT). By transforming the neighborhood of the target triples on the knowledge graph into a line graph, and using the graph attention mechanism to fuse the relationship features on the line graph, this method cannot only reduce the difficulty of modeling the heterogeneous map structure of the knowledge graph, but strengthen the connection between the triples and the entire knowledge graph, thereby reducing the model’s dependence on entity features. The comparison results with related works on the benchmark data sets WN18RR and FB15k-237 prove the superiority of proposed method.