Research On Knowledge Graph Construction Technology Based On Social Network

In recent years,with the rapid development of artificial intelligence technology,knowledge graph research based on data mining,machine learning and knowledge engineering has attracted the attention of many researchers.Knowledge graph has important theoretical and practical significance for the development of artificial intelligence and machine cognitive intelligence.Although researchers have achieved a lot of outstanding research results in knowledge graph construction technology area,many problems still need to be resolved.This dissertation mainly focuses on the research of knowledge graph construction technology based on social network.It mainly studies the problems of entity acquisition,entity disambiguation,relation completion,relation reasoning and knowledge conflict resolution in knowledge graph construction research area.The research results of this paper can provide reference for intelligent applications such as intelligent information retrieval,smart recommendation,and decision support.Specifically,the main innovations achieved in this paper include five aspects.1.As the traditional knowledge graph's data source is relatively limited,three methods are proposed to obtain person entities from social networks.These methods are employed to obtain person entities from social networks according to different social network data features.First,a method based on the topology features of social networks is proposed to associate the users across two networks.This method employs the node's degree centrality,clustering coefficient,and eigenvector centrality as features of random forest classifier to associate the social account across social networks.Second,a method for linking two users across networks is proposed that employs the structural features of social networks and the features of user profiles.This method converts the user's profile feature and social structure feature into feature vectors,and link social accounts according to the cosine similarity of the social account's feature vectors.Third,a method of merging user profiles,user online time distribution features and user interest features is proposed to link users across social network.This method utilizes a bipartite graph matching algorithm to associate social accounts across social networks.Experimental results show that the proposed methods significantly outperform baseline methods.2.A method based on Markov logic network is presented for solving entity disambiguation problem in the process of knowledge graph fusion.In the beginning,the proposed method utilizes knowledge graph to represent the relationship between entities.Then,a reasoning method based on Markov logic network is proposed to inference the conflicts between facts.Lastly,the tensor decomposition method is proposed to disambiguate the entity share the same name.The experimental results show that the proposed method's F1 value,accuracy rate and recall rate are superior to the baseline methods.Among them,experiment results on WD1 and Wikidata show that the proposed method is 8% higher than the DoSeR on F1 value.Experiment results on YA,DB and WD2 datasets show that the average value of F1 value of proposed method is 7% higher than that of DoSeR method.Experiment results on Wikidata and WD1 datasets show that the proposed method's accurate rate is 10% higher than the DoSeR method and the recall rate is 5% higher than the DoSeR method.3.A method named ELPKG is proposed for link prediction between entites in the knowledge graph for solving the problem of missing relationship in the process of constructing knowledge graph.First,ELPKG method utilizes a path-based approach to represent the relationships between entities;then it employs the embedded vectors between the entities to represent the semantic relationships between the entities;then it uses probabilistic soft logic to describe non-deterministic knowledge;and finally proposes a method to combine path-based approach and embedding vectors to complete the entity relationships.The experimental results on real datasets show that the proposed ELPKG method outperforms the existing baseline method in the task of knowledge completion,in which the ELPKG's Hit@1,Hit@10 and MRR are 35%,24%,17% higher than baseline method respectively on the YAGO3 dataset.4.A method named KGIPSL is proposed to solve relation reasoning problem.KGIPSL combines Markov logic network inference model and probabilistic soft logic method for relation reasoning.Fristly,KGIPSL counts the entity relationships;Secondly,KGIPSL utilizes conditional random field method to construct Markov logic network,and KGIPSL employs Local Closed-World Assumption to acquire negative samples;Furthermore,KGIPSL employs a random walk method to sample entities from Markov logic network.Lastly,KGIPSL uses probabilistic soft logic to represent uncertain facts and inference the entities relationship in the process of querying knowledge graph.Experiment results show that the accuracy and efficiency of the KGIPSL method are higher than baseline methods,and the average accuracy of KGIPSL on YAGO was 14.9% higher than the baseline methods.5.A knowledge conflicts resolution method is proposed for solving the problem of knowledge conflicts in the evolution of knowledge graph.This method combines time attribute,semantic embedding representation of the entities and the graph structure into one framework to ensure consistency in the evolution of knowledge graph.First,this method utilizes uncertain temporal knowledge graph to describe the time feature of facts;Then,this method employs time-based conflict feature to add constraint into the facts.Last,a Kcrabdl method based on deep learning is proposed to ensure the consistency of the facts in the evolution of knowledge graph.Results of large number of experiments on IMDB and YAGO datasets show that Kcrabdl method is superior to baseline methods,and it is verified that this method has better robustness with noise data.