Research on reasoning method of multi-relational knowledge graph based on translation

Abstract

In the knowledge reasoning filed, algorithms based on translation have achieved great results in recent years. However, when translation-based models fit multi-map attribute relations, different relations will compete for the same embedding vector. At the same time, many knowledge graphs can automatically extract new triples from the Internet now. This further reduces the frequency of occurrence of a single relation in the knowledge graph, which creates more difficulty for translation-based models. Thus, the TransE_rel model is proposed by relaxing the constraints on relation vectors in the low-dimensional space. This model focuses more on relation reasoning and, when combined with other models, can also be used for entity reasoning. We conduct relation prediction and entity prediction on benchmark dataset and one dataset extracted by ourselves. By comparing with recent work, experimental results show that our method has achieved better performance on two tasks especially on relation prediction.