A Sentence-RCNN embedding model for Knowledge Graph Completion

摘要

Large-scale knowledge graphs are structured to represent real world facts, but they are far from completeness. A number of completion methods have been developed to fill missing facts. In this paper, a novel Sentence-RCNN embedding model is proposed for knowledge graph completion. This model represents knowledge facts as sentences, so that it can precisely capture long-term dependencies, local structure information and translational features simultaneously. In addition, we propose a new method to construct negative samples (CNS), which greatly reduces the number of false negative samples used in model training stage. The proposed model was validated by two challenging benchmark datasets without any extra information. Results showed Sentence-RCNN model had fairly good accuracy, robustness and convergence than the state-of-the-art embedding models. It improved MRR and H©N metrics by an average of over 11.8%, and obtained extra 5.4% improvement with CNS method.