Relation Path Modeling with Entity Types for Knowledge Graph Completion

Abstract

Considering that existing knowledge representation learning methods fail to make full use of various information to enhance knowledge representation, a knowledge representation learning method that incorporates entity types and relation paths is proposed. Firstly, the type-specific projection matrices is constructed by using the hierarchical type information of entities, which allows entities to have different entity representation based on type. Secondly, the representation of relationships between entities is also enhanced by rich semantic information on the path of relationships between entities. Finally, the entity vector and the relation vector are connected to obtain the final knowledge representation. The link prediction task on FB15K dataset shows that PTRL shows significant improvement in MR and Hits@10 compared to mainstream models such as TransE, TKRL and PTransE.