HyperJOIE: Two-View Hyperbolic Knowledge Graph Embedding with Entities and Concepts Jointly

Knowledge graphs have two views: an entity graph in the instance view and a concept graph in the ontology view. Recent studies reveal that modeling the two graphs jointly can benefit the understanding to either one. However, the existing work has flaws on both modelling the hierarchical structures in the Euclidean space, and capturing the deep cross-view interaction between an entity and its corresponding concept. In this paper, we propose to explore hyperbolic space for two-view knowledge graph embedding, which provides more effective and efficient embedding learning mechanism, especially for hierarchical structured knowledge. We also propose to capture the deep cross-view interaction between an entity and its corresponding concept through modeling local structure information from intra-view neighbor nodes with hyperbolic attention mechanism. Finally, we propose to maintain the structural correspondence between the concept graph and the entity graph by first encoding two graphs with the same embedding model respectively and then aligning the two graphs with a hyperbolic transformation. Our empirical study conducted on two benchmark data collections proves that our model outperforms several state-of-the-art two-view knowledge graph embedding models.