Research on Network Link Prediction of Shared Variables Based on Translation Model

In the field of natural language processing, because the discrete symbolized words cannot contain semantic information, mapping words to vector space is not only conducive to the corresponding calculation, but also makes the relevant vectors to contain some semantics in the mapping process. On this basis, the knowledge graph representation represented by deep learning is inspired by the word vector representation, and a series of knowledge representation models are proposed. Although knowledge mappings represented by the the TransE model can solve the problem of data sparsity, it is difficult to distinguish entities with the same semantics when dealing with complex relationships such as one-to-many, many-to-one, and many-to-many. In view of this, this paper proposes an improved shared variable network representation learning method based on translation model, PTransR. The relationship between head and tail entities are represented in their own spaces. In this method, the vector and the relational vector mapped from the head and tail entities to the relational space are regarded as two input embeddings by the method of vector projection distance model, and the prediction problem is regarded as the ordering problem. Finally, an experimental evaluation was carried out on the standard knowledge map public data set FB15K and WN18. The experimental results show that PTransR has a certain improvement in the link prediction task compared with TransR and other models.