Knowledge Graph Embedding with Diversity of Structures

In recent years, different web knowledge graphs, both free and commercial, have been created. Knowledge graphs use relations between entities to describe facts in the world. We engage in embedding a large scale knowledge graph into a continuous vector space. TransE, TransH, TransR and TransD are promising methods proposed in recent years and achieved state-of-the-art predictive performance. In this paper, we discuss that graph structures should be considered in embedding and propose to embed substructures called "one-relation-circle" (ORC) to further improve the performance of the above methods as they are unable to encode ORC substructures. Some complex models are capable of handling ORC structures but sacrifice efficiency in the process. To make a good trade-off between the model capacity and efficiency, we propose a method to decompose ORC substructures by using two vectors to represent the entity as a head or tail entity with the same relation. In this way, we can encode the ORC structure properly when apply it to TransH, TransR and TransD with almost the same model complexity of themselves. We conduct experiments on link prediction with benchmark dataset WordNet. Our experiments show that applying our method improves the results compared with the corresponding original results of TransH, TransR and TransD.