Independent Embedding-Based Relational Enhancement Model for Hyper-Relational Knowledge Graph

The qualifiers (key-value pairs) in the hyper-relational knowledge graphs (HKGs) help the model accurately identify the target. The primary challenge of HKG is how to efficiently obtain the independent features of entities and relations from qualifiers, which can be utilized to address the issue of their potential confusion with the main triples. However, most current models for HKG utilize identical embedding matrices to represent entities within both main triples and qualifiers, making it hard to capture the precise influence of various qualifiers on relations. To address these issues, we propose an improved model IRE for HKG representation. IRE independently learns the embeddings of qualifiers, which promotes exploring interactions and features between qualifiers and main triples. Then, IRE leverages these features to enhance the semantic richness of relations. We incorporate contrastive learning to distinguish entities and relations with unique semantics further, enhancing the model's learning capabilities. IRE can be applied to various downstream tasks, and we conducted experiments using the link prediction task. Experimental evaluations of multiple datasets demonstrate that the IRE consistently outperforms several state-of-the-art baselines.