Research On Learning And Reasoning Over Knowledge Graph

The concept of Knowledge Graph was first proposed by Google.Generally speaking,nodes in Knowledge Graph represent entities of the real world and edges in Knowledge Graph represent relations between entities.Link prediction can predict the probability of a certain relationship between two entities in the Knowledge Graph,for example,it can predict whether a person and a movie or a commodity have a favorite relationship,so as to achieve the purpose of movie recommendation or commodity recommendation;due to v arious reasons,the Knowledge Graph is not complete(for example,common sense knowledge is not added in the construction process,knowledge extraction algorithm is not perfect,RDF triples are lost,etc)When the probability of predicting the relationship R between two entities A and B is high,triple(A,R,B)can be put into the Knowledge Graph,so as to complete the Knowledge Graph.In this paper,the recurrent neural network is used to infer the possible relationship between two entities by using the path between two entities,to complement the Knowledge Graph,and to learn all entities and relationships into emmbeddings at the same time.The main contributions of this paper are as follows: 1.The experimental results are compared with those of link prediction that only focus on the relationship on the path and at the same time focus on the relationship and entity type on the path;2.Introduce a simple attention mechanism to solve the problem of too many useless paths between entities and the problem of training speed;3.Do experiments to compare the effect of different similarity measurement functions on the link prediction of recurrent neural network influence.We do experiments on Free Base dataset,and the results show that considering the entities and relationships on the path,choos ing cosine distance for similarity measurement function,and examin ing the path vectors similar to the top 3 of prediction relationship vector perform best.The m AP(mean average precision)reaches 0.731,which is bett er than the typical algorithms PRA(Path Ranking Algorithm)and PTranse in link prediction task.