HKNI: Fusing Neighbor Information for Hyper-Relational Knowledge Graph Completion

The hyper-relation knowledge graph constructed by hyper-relation facts can effectively express most of the information in the real world. Similar to traditional knowledge graphs, it also has more or less missing information, which makes link prediction on the hyper-relation knowledge graph equally important.However, due to the extension from binary relations to n-ary relations, the hyper-relation knowledge graph has both graph representations based on hypergraphs and sequential representations based on semantics.Most existing models mainly focus on handling only one type of representation within the internal structure of hyper-relations, which limits the comprehensive learning of hyper-relation knowledge graphs.Moreover, there is a large amount of neighbor information in hyper-relation knowledge graphs, but most existing models simply feed the sequential representations into Transformer layers, which greatly reduces the model’s ability to handle neighbors.Therefore, based on the two aforementioned issues, this paper presents a Hyper-Relation Knowledge Graph Completion model that integrates neighbor information(HKNI). HKNI not only effectively handles the two types of representations within the hyper-relation, but also efficiently processes neighbor information.In the end, we compared our model with existing baseline models and experimental results demonstrated that our model achieves the best performance in link prediction tasks.