Multi-perspective semantic decoupling and enhancement in graph attention network for knowledge graph completion

Tianyi Xu, Yan Wang, Wenbin Zhang, Yue Zhao, Jian Yu, Mei Yu, Jiujiang Guo, Mankun Zhao

Abstract: Knowledge Graphs (KGs) are semantic repositories that describe the real world and have been widely applied in various downstream applications. However, KGs still have many incomplete facts, so Knowledge Graph Completion (KGC) is proposed to infer missing facts. Among them, Graph Attention Network-based models (GATs) show great power. However, GATs have two flaws in handling multiple semantics of entities in relational context: (1) Current GATs fail to distinguish the various semantics of the entity which are exhibited by the relations from different perspectives. (2) Existing GATs cannot capture the similar semantics of different entities which are presented by the relations from the same perspective. Hence, we propose a graph attention network based on multi-perspective semantic decoupling and enhancement (MSDE). To capture diverse semantics in the relational context, we classify relations to partition entity multi-perspective semantics, and then we use graph attention networks to obtain multi-perspective decoupled embeddings of entities. To capture semantically similar entities, we select multi-perspective similar entities based on multi-perspective conditional entropy and high-order similar neighbors based on multi-perspective decoupled embedding. Finally, we use an attention decay network to aggregate multi-perspective similar entities and high-order similar neighbors to update entity feature embeddings. Experimental results show that MSDE exhibits marked performance gains compared to other state-of-the-art (sota) models. Significantly, the MRR indicator improves by 6.5% on the FB15K-237 dataset, by 2.3% on the WN18RR dataset, by 7.3% on the Kinship dataset and by 9.2% on the YAGO3-10 over the sota models.