Pre-train Unified Knowledge Graph Embedding with Ontology

Existing knowledge graph embedding models mainly focus on a single task, such as link prediction or entity typing, which actually cannot ensure the generalization capability of the model. Recent research shows that introducing additional ontology information can naturally convert the entity typing task to a specific case of link prediction between the instance and ontology layers. However, the unbalanced scale of the two layers brings difficulty for learning. To this end, we pre-train the knowledge graph embedding on the instance and schema layers of KG respectively on the basis of Rot-Pro, a model that is capable to express the transitivity relation pattern occurred in class hierarchy of the ontology. Furthermore, we construct a dataset by integrating entity type and class hierarchy information based on YAGO3 for evaluating the model efficiency on both link prediction and entity typing tasks. Experimental result shows that our model provided a unified and effective approach for both tasks.