Separate Spaces Framework for Entities and Relations Embeddings in Knowledge Graph

Abstract

Learning Entity and Relation Embeddings is to represent the entities and relations between them in knowledge graph. Recently a variety of models, starting from the work of TransE, to a series of following work, such as TransH, TransR, TransD, are proposed. These models take a relation as transition from head entity to tail entity in principle. The further researches noticed that relations and entities might be able to have different representation to be cast into real world relations. Thus it improved the embedding accuracy of embeddings. Based on these works, we further noticed that in some real world cases entities representation could be vary, especially when one entity in different positions of a relation. In this paper we proposed a new relation embedding by distinguishing head entity and tail entity. It is possible that an entity can have different semantic meanings in head and tail positions. Motivated by this fact, we defined two different entity spaces, head entities and tail entities, for each type of entities. With the training of the general embeddings learning model, we projected the two entity spaces into relation space to run the stochastic gradient descent (SGD) optimization. Experimental results show our method can have good performance in typical knowledge graphs tasks, such as link prediction.