H2KGAT: Hierarchical Hyperbolic Knowledge Graph Attention Network
Knowledge Graphs encode rich relationships among large number of entities. Embedding entities and relations in low-dimensional space has shed light on representing knowledge graphs and reasoning over them, e.g., predicting missing relations between pairs of entities. Existing knowledge graph embedding approaches concentrate on modeling symmetry/asymmetry, inversion, and composition typed relations but overlook the hierarchical nature of relations. Recent studies have observed that there exist rich semantic hierarchical relations in knowledge graphs such as WordNet, where synsets are linked together in a hierarchy. To fill this gap, in this paper, we propose Hierarchical Hyperbolic Knowledge Graph Attention Network (H2KGAT), a novel knowledge graph embedding framework, which is able to better model and infer hierarchical relation patterns. Specifically, H2KGAT defines each entity in a hyperbolic polar embedding space. In addition, we propose an attentional neural context aggregator to enhance embedding learning, which can adaptively integrate the relational context. Our empirical study offers insights into the efficacy of modeling the semantic hierarchies in knowledge graphs, and we achieve significant performance gains compared to existing state-of-the-art methods on benchmark datasets for link prediction task, particularly at low dimensionality.
PDF Abstract
