An Ontology-enhanced Knowledge Graph Embedding Method
Changlong Wang, Tingting Gan, Xingyu Li, Linghan Zhang, Xijie Wang · 2023
Knowledge graph embedding maps entities and relationships of the graph into low dimensional dense vectors to expresse the semantic information, meantime provides effective support for downstream tasks such as link prediction. However, the existing knowledge graph embedding methods mainly focus on the explicit structured information in the graph and rarely use the entailed rich ontological knowledge. Therefore in the paper, a method for injecting ontology information into the embedding model is proposed, ontology information including class hierarchy information and relationship attribute constraints,especially symmetry attributes are considerd. By taking ontology information as extra constraints, the loss function is further refined.the generation of training samples is optimized and the number of false negative samples is limited. Experiments on the two datasets of DBpedia15K and NELL show that the embedding model can be further optimized by injecting ontology information. Specially, the hit rate of triple prediction is 63.70% for the no-type, and for type-triples the MR and the H@10 are 13.51 and 96.59% respectively. The proposed model has better performance than the basic model, which further confirms the effectiveness of the prior knowledge of ontology in knowledge graph embedding learning.