Knowledge Graph Completion via Bidirectional Translation and Interaction

Knowledge graphs have achieved great success for many artificial intelligence related downstream tasks. However, they are still far from being incomplete. In order to automatically predict missing facts, researchers proposed many knowledge graph completion methods. To address the issues that translation-based methods in handling the complex relations and symmetric relations of knowledge graph completion task, a new knowledge graph completion method via bidirectional translation and interaction called BI-TransE is proposed in this paper. First, BI-TransE embeds the entities and relations of triples into low dimensional vectors. Then, BI-TransE models the complex relations, such as 1-to-N, N-to-1, N-to-N, through the interaction between entities and relations, and models symmetric and inverse relations leveraging bidirectional translation score function. We evaluate our BI-TransE for knowledge graph link prediction task. Experimental results show that, BI-TransE can work well on complex relations and symmetric and inverse relation connectivity patterns, and achieves new state-of-the-art performance on four large-scale popular datasets.