Abstract: Recently,representation learning technology represented by deep learning has attracted many attentions in natural language processing,computer vision and speech recognition.Representation learning aims to project the interested objects into a low-dimensional,dense and real-valued semantic space.To this end,a number of models and methods were proposed for knowledge embedding.Among them,TransE is a classic translation-based method with low model complexity,high computational efficiency and favorable knowledge representation ability.However,it has limitations in dealing with complex relations including reflexive,one-to-many,many-to-one and many-to-many relations.In light of this,this paper proposed a revised translation-based method for knowledge graph representation,namely STransH.In this method,firstly,entity and relation embeddings are built in separate entity space and relation space,and then the non-linear operation of single-layer network layer is adopted to enhance the semantic connection between entity and relation.Inspired by TransH,this paper introduced the relation-oriented hyperspace model,thus projecting head and tail entities to the hyperspace of a given relation for distinction.Besides,it also proposed a simple trick to improve the quality of negative triplets.At last,it conducted extensive experiments on link prediction and triplet classification on benchmark datasets like WordNet and Freebase.Experimental results show that STransH performs significant improvements over TransE and TransH compared with TransE and TransH,and its Hits@10 and triplet classification accuracy are increased by nearly 10% respectively.

Key words: Knowledge graph, Representation learning, Link prediction, Triplet classification