Learning knowledge graph embedding with a dual-attention embedding network. (February 2023)

• Record Type: Journal Article
• Authors: Fang, Haichuan; Wang, Youwei; Tian, Zhen; Ye, Yangdong
• Is Part Of: Expert systems with applications. Volume 212(2023)
• Publication Date: 2023-02
• Subjects: Knowledge graph embedding -- Knowledge graph -- Graph convolutional network -- Representation learning -- Attention mechanism; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118806
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Knowledge Graph Embedding (KGE) aims to retain the intrinsic structural information of knowledge graphs (KGs) via representation learning, which is critical for various downstream tasks including personalized recommendations, intelligent search, and relation extraction. The graph convolutional network (GCN), due to its remarkable performance in modeling graph data, has recently been studied extensively in the KGE field. However, when learning entity representations, most attention-based GCN approaches treat neighborhoods as a whole to measure their importance without considering the direction information of relations. Additionally, these approaches make relation representations perform self-update via a learnable matrix, resulting in ignoring the impact of neighborhood information on representation learning of relations. To this end, this study presents an innovative framework, namely learning knowledge graph embedding with a dual-attention embedding network (D-AEN), to jointly propagate and update the representations of both relations and entities via fusing neighborhood information. Here the dual attentions consist of a bidirectional attention mechanism and a relation-specific attention mechanism for jointly measuring the importance of neighborhoods in respectively learning entity and relation representations. Thus D-AEN enables elements like relations and entities to interact well semantically, which makes their learned representations retain more effective information of KGs. Extensive experimental results on three standard link prediction datasets demonstrate the superiority of D-AEN over several state-of-the-art approaches.