Enhancing Semantic Awareness in Knowledge Graph Embedding

Knowledge Graph Embedding aims to represent entities and relations as low dimensional vectors. The representative ability of low dimensional vectors is one of the most important aspects of knowledge graph embedding. Knowledge graph embedding needs to represent two kinds of information: entity relation structures and entity semantic differences. However, the most existing approaches are good at entity relation structure representation, but fail to represent entity semantic differences which are common in real-world applications. To address the challenge, we propose a novel approach called SeTransE to enhance semantic awareness in knowledge graph embedding and achieve the aim of representing entity semantic differences. It mainly adopts a new method where a relation can rotate to a certain angle, which is different from traditional approaches. Experimental results on multiple benchmark knowledge graphs show our proposed approach not only represents relation structures, but also effectively models semantic structures and distinguish semantic differences.