Jointly Modeling Structural and Textual Representation for Knowledge Graph Completion in Zero-Shot Scenario

摘要

Knowledge graph completion (KGC) aims at predicting missing information for knowledge graphs. Most methods rely on the structural information of entities in knowledge graphs (In-KG), thus they cannot handle KGC in zero-shot scenario that involves Out-of-KG entities, which are novel to existing knowledge graphs with only textual information. Though some methods represent KG with textual information, the correlations built between In-KG entities and Out-of-KG entities are still weak. In this paper, we propose a joint model that integrates structural information and textual information to characterize effective correlations between In-KG entities and Out-of-KG entities. Specifically, we construct a new structural feature space and build combination structural representations for entities through their most similar base entities. Meanwhile, we utilize bidirectional gated recurrent unit network to build textual representations for entities from their descriptions. Extensive experiments show that our models have good expansibility and outperform state-of-the-art methods on entity prediction and relation prediction.