Knowledge Graph Completion via Complete Attention between Knowledge Graph and Entity Descriptions

The objective of learning representation of knowledge graph is assumed to encode both entities and relations into a continuous low-dimensional vector space. Previous methods usually represent the same entity in different triples with the same representation. Considering the fact that different entities should have different semantics in different triples, this paper proposes a method to learn entity description information based on the triple, under the complete attention (CATT) mechanism, in the knowledge graphs. By doing so, the entity has different representations of corresponding semantics in different triples. For encoding the information of entity description, we use three deep learning methods including CNN, Bi-LSTM and Transformer. Experimental results show that, with the proposed method, the performance on both entity prediction and relationship prediction improved significantly.