Knowledge Graph Completion with Fused Factual and Commonsense Information

Changsen Liu, Jiuyang Tang, Weixin Zeng, Jibing Wu, Hongbin Huang

Abstract: Knowledge graph contains a large number of factual triples, but it is still inevitably incomplete. Previous commonsense-based knowledge graph completion models used concepts to replace entities in triplets to generate high-quality negative sampling and link prediction from the perspective of joint facts and commonsense. However, they did not consider the importance of concepts and their correlation with relationships, resulting in a lot of noise and ample room for improvement. To address this problem, we designed commonsense for knowledge graph completion and filtered the commonsense knowledge based on the analytic hierarchy process. The obtained commonsense can further improve the quality of negative samples and the effectiveness of link prediction. Experimental results on four datasets of the knowledge graph completion (KGC) task show that our method can improve the performance of the original knowledge graph embedding (KGE) model.

External IDs:dblp:conf/IEEEwisa/LiuTZWH23