Sememes-Based Framework for Knowledge Graph Embedding with Comprehensive-Information

Abstract

The goal of knowledge graph embedding is to represent both entities and relationships as low-dimensional, dense vectors that can be used to empower other machine learning models. While most approaches concentrate on modeling the structural information of the graph, part of the work also focuses on fusing entity descriptions, allowing entities to be fused with richer semantics. However, the complex entity text descriptions contain a lot of noise, which reduces the semantic purity. Therefore, in this paper, we propose a novel sememes-based framework for knowledge graph to streamline the semantic space of entities. More specifically, We replace entity descriptions with a finite set of semantics and encode the sememe labels of entities using a pre-trained Bert model, and finally jointly learning the symbolic triples and sememe labels. The experimental results show that our method outperforms other baselines on the task of link prediction and entity classification.