Study On Knowledge Graph Completion Algorithm Based On Knowledge Representation Learning

Knowledge graph describes the concepts,entities and their relationships in the objective world in the form of triples,which provides a better expression for automatic understanding of massive information.In the knowledge map constructed by using large-scale knowledge acquisition method,there are usually a large number of relationship missing problems,such as 71% of entities in freebase are lack of birth place related information,Word Net and Nell also have different degrees of part of speech,parents and other relationship missing.The missing relationship will reduce the accuracy of downstream application tasks.Knowledge graph completion is an automatic completion method for this problem,mainly based on symbol reasoning,link prediction and knowledge representation learning.The method based on symbolic reasoning mainly relies on artificial prior rules,which is inefficient and cannot be applied to large-scale knowledge graph.The methods based on link prediction use the network topology,assuming that similar nodes are most likely to form links,but links in the knowledge graph have clear semantic information,which makes the link prediction effect not ideal.The completion of knowledge representation learning is a method combining representation learning and link prediction.It uses numerical reasoning instead of symbolic reasoning,and provides an efficient way for the completion of large-scale knowledge base.In this paper,three aspects of research have been carried out to solve the problems of computational efficiency and structural information in knowledge representation learning.(1)Considering the problem of single node relationship and low computational efficiency in link prediction in the existing network representation learning.We use the "resource allocation" of common neighbors between nodes in the network to characterize the similarity between nodes.The similar nodes are mapped into similar vector spaces to learn the representation of the nodes.A link prediction network representation learning method(NELP)based on resource allocation is proposed.(2)Aiming at the problem that the existing knowledge representation learning model is becoming more and more complicated and has too many parameters.The relationship type score function is defined by using the relationship type constraint information in the knowledge graph,and the score function is combined with the Trans E score function.A knowledge representation learning method(Trans RT)based on constraint type prior information is proposed.(3)Aiming at the problem that the existing knowledge representation learning model has a single score function and insufficient consideration of structured information.A structured probability score function 1)(?,,)is proposed using the probability directed graph model.And we proposed a knowledge representation learning method(PGME)based on probability graph model.