Cross-Projection for Embedding Translation in Knowledge Graph Completion

Abstract

Knowledge graphs have been applied in many applications. However, they are incomplete. Previous completion works such as TransE, TransH and TransR/CTransR regard relation as translation between head entity and tail entity. They think that only entities are affected by relations, and change entities embeddings according to the corresponding relation. However, they neglect the influence between two entities. In this paper, we propose a new model named Cross-projected Translation Embedding model (CpTE) based on the translation theory. In CpTE, we assume that relation is affected by related entities and transmits effect between end-to-end entities. Head entity and tail entity are influenced by each other rather than their relation. Given a triple, we project each entity into another entity space and map relation into a union space constructed by entities. In Experiments, we evaluate our model on two typical tasks including link prediction and triple classification. Experiments results show that our approach outperforms than TransE, TransH and TransR/CTransR in link prediction task on evaluation sets WN18 and FB15K, and gets an improvement in triple classification on specific evaluation set FB15K.