Research On Knowledge Graph Completion Algorithm Of Translation-based Model

The knowledge graph extracts knowledge from the Internet information of the Internet and organizes it into structured data in the form of triples,which is convenient for computer to process efficiently.It is widely used in information retrieval,intelligent recommendation,question and answer systems and other fields.A large number of knowledge maps have been constructed in the current large-scale knowledge map,but they are still incomplete,and there is a lot of information missing.The completion algorithm for perfecting the knowledge map has become a research hotspot.Knowledge representation learning is an efficient means to realize the completion of knowledge graph.It expresses the entities and relationships in the knowledge graph as low-dimensional dense vectors,and infers the potential semantic relations by vector calculation,which greatly reduces the computational complexity.In recent years,a large number of researches on knowledge representation learning have emerged.Among them,the TransE is the most representative.It pioneered the rules of translation,simplified the connection between entities and relationships,and significantly improved the effect of knowledge graph completion.The TransE model and its extended model are analyzed in detail.In view of the fact that TransE is difficult to deal with complex relationships,the WTransE model is designed to deal with triples in a weighted manner,so that entities can express different semantics in different translation operations.At the same time,also integrates the entity similarity into the model,ensuring that the embedded vector retains the original similarity feature better and improves the representation ability of the embedded vector.Entity similarity is considered from two aspects,one is the similarity feature mined from the triple structure,and the other is the similarity according to the text description of the entity.Experiments on knowledge graph completion tasks such as link prediction andtriple classification on the standard dataset for the proposed model and method.The experiment confirmed that the performance of WTransE model is greatly improved compared with TransE,and the entity similarity is integrated.The method can further improve the performance of the model,indicating that the model and method are effective and feasible.