Knowledge Graph Completion Method of Combining Structural Information with Semantic Information

With the development of knowledge graphs, a series of applications based on knowledge graphs have emerged. The incompleteness of knowledge graphs makes the effect of the downstream applications affected by the quality of the knowledge graphs. To improve the quality of knowledge graphs, translation-based graph embeddings such as TransE, learn structural information by representing triples as low-dimensional dense vectors. However, it is difficult to generalize to the unseen entities that are not observed during training but appear during testing. Other methods use the powerful representational ability of pre-trained language models to learn entity descriptions and contextual representation of triples. Although they are robust to incompleteness, they need to calculate the score of all candidate entities for each triple during inference. We consider combining two models to enhance the robustness of unseen entities by semantic information, and prevent combined explosion by reducing inference overhead through structured information. We use a pre-training language model to code triples and learn the semantic information within them, and use a hyperbolic space-based distance model to learn structural information, then integrate the two types of information together. We evaluate our model by performing link prediction experiments on standard datasets. The experimental results show that our model achieves better performances than state-of-the-art methods on two standard datasets.