Joint Global-Local Representations via Relation-Entity Pair Encoding for Hyper-Relational Knowledge Graphs

Hyper-relational knowledge graphs (HKGs) extend knowledge graphs with qualifiers to represent complex n-ary facts. However, existing methods often (1) rely on relation-agnostic node aggregation that mixes heterogeneous multi-hop evidence, which can lead to entity-role ambiguity and representation collapse, and (2) fail to jointly model global inter-fact structure and local intra-fact semantics, limiting mutual refinement during training. We propose GLoRE (Global-Local Representations via Relation--Entity pair encoding), a joint global-local model that treats relation--entity pairs as the modeling primitive to preserve relation identity and within-fact roles, thereby retaining role-specific signals (i.e., reducing role mixing) across multi-hop aggregation. To provide a relation affinity prior based on co-occurrence statistics, we construct a weighted relation graph from relations that co-occur in the same fact and use it to refine relation representations in the pair encoder. On top of pair embeddings, GLoRE performs global hypergraph propagation over pairs to capture cross-fact dependencies, and local intra-fact self-attention over pair sequences to model fine-grained semantics within each fact. The two views are optimized jointly end-to-end with shared pair encoding, enabling training-time global--local coupling. Experiments on JF17K, WikiPeople-, and WD50K achieve state-of-the-art link prediction performance, with 5-25% relative improvements in MRR and ranked-first results across all qualifier densities, demonstrating robustness to qualifier sparsity. Our code is available at https://github.com/Approxy02/GLoRE.