Representation Learning of Knowledge Graph Integrating Entity Description and Language Morphological Structure Information

Knowledge graph embedding, which projects the symbolic relations and entities onto low-dimension continuous spaces, is the key to knowledge graph completion. The representation learning methods based on translation, such as TransE, TransH and TransR, only consider the triple information of knowledge graph, and fail to make effective use of other information of entity. To solve these problems, in this paper, we propose a knowledge graph representation learning method which integrates entity description and language morphological structure information to deal with complex relations (i.e. 1-N, N-1 and N-N relations). Firstly, the fastText model which considers affix of words is used to get the embedding of all entity description information. Then, the triple embedding, entity description embedding are spliced to obtain the representation of the final entity embedding. In addition, we propose a new score function-distcos–man, which considers the similarity of entity vector not only from the value of each dimension, but also from the direction of vectors. Experiments show that our method achieves substantial improvements against state-of-the-art baselines, especially the Hit@10s of head entity prediction for N-1 relations and tail entity prediction for 1-N relations improved by about 11.6% and 17.9% on FB15K database respectively.