Potential probability of negative triples in knowledge graph embedding

Abstract

In this paper, we propose the concept of triples’ potential probability. Typically, knowledge graph only contains positive triples. Most of knowledge representation methods treat the replaced triples, which replace the head/tail entities or relations with other entities or relations randomly, as negative triples. Actually, not all triples are absolutely negative triples after substitution. It could be a positive triple essentially, but has not been discovered yet. Considering the problems arising from the above situation, we propose the potential probability to solve it. First, we utilize the co-occurrence of relations and paths in the knowledge graph to find potentially correct probabilities of some negative triples. Then we add these triples with potential probabilities to the training model. Finally, we take the experiments on two translation-based models, TransE and TransH, using four public datasets. Experimental results show that our method greatly enhances the performance of the target embedding models.