Research On Knowledge Graph Completion Technology Based On Representation Learning

As the core of cognitive intelligence,Knowledge Graph(KG)has been successfully applied in intelligent search,personalized recommendation,and other fields.However,there are still many quality problems in the current KG.Incomplete knowledge is one of the important problems that restrict the effectiveness of the KG in the application field.This paper mainly studies the knowledge graph completion method based on representation learning in order to deal with this problem.In the beginning,this paper summarizes the evolution of KG in recent years,explains KG's quality issues,and highlights the research context and significance.Then,this paper discusses the current state of KG research at home and abroad,and the relevant research foundations and technologies,as well as some issues with current KG completion methods.This paper proposes two KG completion methods based on representation learning:(1)This paper combines the attention mechanism with the embedding representation of knowledge and proposes a completion method for KG named Pattention.Because the correlation between the entity and different neighborhoods is different.the attention mechanism is introduced here to assigned different weights to different neighborhood information when Pattention embed the knowledge graph.It can fuse the neighborhood information to embed the entity.In this method,a convolutional neural network is introduced to obtain the characteristics of entities and relationships,and a bidirectional long-term and short-term memory network are introduced to translate the path into embedded vector.After introducing the path information,this method uses the attention function to score the path information,and when combined with the score,the final state vector of the candidate triple is obtained.Finally,the probability score is made for the triple where the candidate entity is located.This paper tests and compares the method with the relevant baseline methods.In the experiment on NELL995,the completion effect of this method is improved by 8% in the MRR index.The experimental results show that Pattention performs better than the relevant methods.(2)This paper combines the characteristics of KG and attention mechanism to put forward an graph attention faded mechanism,and designs a knowledge graph completion method on this basis named GAFM.GAFM combines the path length information in the KG with the attention value.The purpose is to obtain a more expressive representation of the target entity by fusing the neighborhood information of the target entity into the representation of the target entity.This method introduces a capsule neural network model to realize feature extraction.In this paper,GAFM is run in the link prediction experiment and compared with the baseline method.The results of this method on FB15K-237 show that GAFM has been improved by 7%and 8% in Hits@3 and Hits@10 compared with the second-ranking method.And GAFM achieved better knowledge graph completion effect.At the same time,this paper constructs a computer major KG based on the proposed two completion methods and the related technologies of KG.We apply the KG completion methods proposed in this paper to the computer major KG to improve the problem of incomplete knowledge in the KG.We have also built an online learning system based on the computer major KG.This paper shows the relevant functions of the system.