Concept-driven representation learning model for knowledge graph completion

Yan Xiang

doi:10.1016/j.eswa.2024.126297

Concept-driven representation learning model for knowledge graph completion

2025-04-01

0

PRE

AI

DOI：10.1016/j.eswa.2024.126297

Original

Original request for help

Save

Abstract

En 中文

Knowledge graph completion (KGC) aims to address the problem of incomplete knowledge graph (KG) by predicting missing entities or relations. Representation learning-based KGC methods have shown good performance. However, current research mainly focuses on learning semantic representations of individual entities, often overlooking the role of entity sets. To address this, we propose a new representation learning model that enhances the representation ability for entities by leveraging both intra-correlations and inter- correlations of entity sets. Specifically, we consider sets of entities constrained by specific relations as concepts, and construct homo-concept and hetero-concept entity sets. We then select positive and negative sample pairs from these different concept sets for contrastive learning to obtain basic representations of entities. Building on this, we further refine the comprehensive representations of entities by utilizing the similarities between entities and the set center within the same homo-concept set, as well as the similarities between two homoconcept sets with specific relations. Experiments show that the comprehensive representations learned by our model achieve accurate results in KGC tasks across five benchmark datasets. Compared to other state-of-the-art models, our model demonstrates significant advantages.

Keywords:

Knowledge graph

Knowledge graph completion

Contrastive learning

Representation learning

Concept-driven representation learning model for knowledge graph completion

Abstract

Journal

Organization