XLNet For Knowledge Graph Completion

Abstract

As the one of databases in artificial intelligence systems, knowledge graph has been widely used nowadays. Although the number of entities in current knowledge graphs has reached tens of millions or even billions level, directed cyclic graphs of their relations and entities composition were still relatively sparse. Knowledge graph completion can make the structure and content of the knowledge graphs completer and richer. In this paper, the knowledge graph completion was converted into classification and scoring tasks, and several knowledge graph completion models based on XLNet were proposed. Besides, a new negative sampling method was also proposed to improve the quality of the negative samples. Several comparative experiments were done for different output layers. Experiments on several benchmark knowledge graphs shown that our methods achieved state-of-the-art performance in common evaluation methods for knowledge graph completion.