Improved model for knowledge representation: TransMR

Abstract

Knowledge graphs have developed rapidly in recent years, and knowledge representation learning is a fundamental task of knowledge graphs, so knowledge representation learning has likewise received widespread attention. Therefore, a series of knowledge representation models based on TransE methods have been proposed by scholars one after another, among which, the translation model TransE has low model complexity, high computational efficiency, and strong semantic representation ability for knowledge representation of triples. However, the TransE method cannot handle dealing with complex relations. In view of this, this paper proposes an improved knowledge representation model TransMR based on the TransE method, which uses the Marxian distance instead of Euclidean distance to calculate the distance between vectors, and builds entity and relationship models in entity space and relationship space respectively, in which the back propagation and nonlinear operation of single layer neural network are used to enhance the semantic connection between them. Meanwhile, during the model training process, experiments are conducted to improve the fault tolerance of negative example triples by using substitution for the most similar entities.