Dual-context enhanced knowledge representation learning method in hyper-relational knowledge graphs

Jiahang Li, Qilong Han, Dan Lu, Lijie Li

Abstract: In hyper-relational knowledge graphs (HKGs), learning effective knowledge representation using qualifiers improves the accuracy of tasks. Most methods use shared embeddings for qualifiers and triples that ignore their distinct roles and cause semantic confusion, while the unique structure of HKGs introduces complex contexts that make it a challenge to explore semantics deeply. Therefore, providing independent representations and effectively integrating contextual information is crucial. The Dual-Context Enhanced Knowledge Representation Learning (DC-IRE) model is proposed to address these challenges. First, multiple independent vectors for entities and relations are learned according to their specific roles within triples and qualifiers, ensuring that the representations adapt to different contexts and eliminate potential semantic confusion. Then, features extracted from various contexts in HKGs are aligned and integrated to improve the simplistic update process for relations, as relations are crucial bridges connecting the entire graph. For this purpose, HKGs are transformed into relation subgraphs, which facilitates extracting semantic and entity type contextual features. This operation addresses the issue that most methods neglect the learning of relation representations, with these methods typically updating using a parameter matrix without considering the crucial information within HKGs. Finally, the distinction of entities and relations with unique semantics is achieved by employing a relation enhancement module and multi-task optimization, thereby enhancing the learning capability of the model. The effectiveness of the DC-IRE model is validated through the link prediction task, with experiments on multiple datasets showing an improvement of up to 13% in MRR over baseline models.

External IDs:dblp:journals/comintsys/LiHLL25