Improvement on message passing of hyper-relational knowledge graph

Abstract

Unlike traditional knowledge graphs (KGs), which represent real-world facts as entity-relationship-entity triples, hyper-relational knowledge graphs allow triples of traditional entities with additional relation-entity pairs (also known as qualifiers) to establish connections to deliver more complicated messages. Modelling triplet-qualifier relationships effectively for hyper-relational knowledge graphs to accomplishing prediction tasks is an existing challenge. In this paper, an improved hyper-relational knowledge graph completion method, STARE, is proposed by introducing two new methods: (1) Reconstructing the graph neural network module in the original STARE; (2) Introducing centralization and scaling to the model-the idea of shrinking to avoid over-smoothing. In our experiments on four benchmark datasets, the proposed method is always better than STARE, improving the prediction ability of the completion of HKGs.