Knowledge Graph Completion with Triple Structure and Text Representation

Abstract

Abstract Knowledge Graphs (KGs) describe objective facts in the form of RDF triples, each triple contains sufficient semantic information and triple structure information. Knowledge Graph Completion (KGC) is to acquire new knowledge by predicting hidden relationships between entities and adding the new knowledge to the KG. At present, the mainstream KGC approaches only applied the triple structure information or only utilized the semantic information of the text. This paper proposes an approach (TSTR) using BERT and deep neural networks to fully extract the semantic information of knowledge, and designs an aggregated re-ranking scheme that incorporates existing graph embedding approach to learn the structural information of triples. In experiments, the approach achieves state-of-the-art performance on three benchmark datasets, and outperforms recent KGC approaches on sparsely connected datasets.

Keywords

• Knowledge graph completion
• BERT
• Deep convolutional architecture
• Re-ranking