KGESS - A Knowledge Graph Embedding Method Based on Semantics and Structure

To achieve a better performance in the downstream task of knowledge graph (KG), a good representation of KG is necessary. Sensing from the topological structure of the graph, most conventional methods tend to ignore the semantic features of nodes, which is significant for describing the entity in KG. In this paper, we propose a novel Knowledge Graph Embedding method based on Semantics and Structure (KGESS), which learned the representation of KG from both topological facts and semantic information. It leverages Chinese BERT to obtain semantic features of the entity first. Then it further enhances these features via a neural module, namely Semantic Feature Extractor. To evaluate the performance of KGESS, we utilize an additional linear module to execute the link prediction task. Experimental results demonstrate that KGESS achieves a superior Hit@k score than conventional methods, indicating the effectiveness of the idea of enhancing structure with semantics in the representation task of KG.