Knowledge graph (KG) plays an important role in many downstream tasks of information retrieval and natural language processing, as it provides a way of utilizing common knowledge. However, due to its symbolic nature, it is difficult to manipulate information in it. Knowledge graph completion models alleviate this problem by embedding the facts from a knowledge graph into continuous vector space, i.e. the feature space, converting manipulations of the information to operations in vector space so that one can utilize the information in a flexible manner.
One of the critical tasks of knowledge graph completion model is link prediction. For a triplet (h,r,?) with one term t missed, the model evaluates the plausibility of each entity to be the candidate for t. However, most models do not consider type constraints of relations in this process. We proposed an unsupervised method to infer type constraints by the geometric distribution of knowledge graph embeddings. Results on three widely used datasets show that our method can improve the performance of knowledge graph completion models with fewer parameters.