Representation purification with semantic contrastive learning for hyper-relational knowledge graph

Jianqiang Huang

doi:10.1016/j.knosys.2025.114898

Representation purification with semantic contrastive learning for hyper-relational knowledge graph

2025-11-14

0

PRE

AI

DOI：10.1016/j.knosys.2025.114898

Original

Original request for help

Save

Abstract

En 中文

• The importance and originality of this study lie in its exploration of representation learning based on real-world HKGs with low-quality annotations. Our study provides new insights into enhancing H-Facts representation and HKGs by incorporating hypergraph-aware and sequence-aware semantic features. • To our knowledge, HKGCL represents the first model to apply contrastive learning to representation learning for HKGs. This approach involves training contrastive semantic representations to identify entities with strong correlations by purifying their features. Furthermore, we present a theoretical analysis to support the benefits conferred by these integrative methods. • We address the limitation of using non-interactive semantics to represent H-Facts by proposing novel learnable residual connections (LRCs) that dynamically learn semantic representations based on hypergraph-aware and sequence-aware semantic interactions. Incorporating LRC allows the model to reduce noise interference and enhance the semantic representation of H-Facts. • Beyond link prediction, we validate cross-task generalization on entity alignment and qualifier-aware relation classification derived from the same HKGs, and outline a realistic e-commerce deployment scenario to illustrate practical usability.

Representation purification with semantic contrastive learning for hyper-relational knowledge graph

Abstract

Journal

Organization