Research On Knowledge Graph Fusion Construction And Completion Method

With the development of network interconnection,the data on the Internet grows exponentially,and most information is stored in the form of database tables.The knowledge graph can effectively organize,store,and represent these database table data.Knowledge graph is composed of a large number of entities and relationships,with strong semantic expression and data modeling ability,and is widely used in various fields such as automatic question and answer,search engine,personalized recommendation and so on.However,there are still some problems in the construction and fusion of current knowledge graphs based on database data.Typically,only datasets with similar properties can be fused.Moreover,duplicate nodes may occur when building a knowledge graph,leading to problems such as storage redundancy.At the same time,most of the existing knowledge graph are built manually or semi-automatically,resulting in the absence of a large number of entities and relationships in the knowledge graph,resulting in incomplete and sparse data of the knowledge graph.Therefore,it is urgent to use knowledge graph completion methods,that is,link prediction tasks to complete the missing knowledge in the knowledge graph.To solve the above problems,the following research is conducted:A method of knowledge graph fusion and construction based on database tables is proposed.In view of the problem of node redundancy caused by the generation of database table knowledge graph,the lack of knowledge of single table and the difficult mapping of multi-source database tables without relationship constraints,this thesis proposes the fusion construction method of knowledge graph based on database tables.This method first uses the TKGC method proposed in this thesis to select the core attribute value,attribute name and attribute value> ternary table knowledge graph,and then use the SNF fusion method or SRF fusion method of different types of database table knowledge graph,and finally realize the visual storage based on Neo4 j.In this thesis,four real data sets are used for simulation experiments,and the visualization results prove that the constructed graph is true and effective.After the fusion method,the overall node ratio of the knowledge graph increases by22.3% and the number of relationships increased by 10.5%,which proves that the method of this chapter enhances the ability of graph joint query and knowledge mining.A link prediction method(NALP)based on neighborhood information and attention mechanism is proposed.In link prediction,most of the existing knowledge completion models are independent triples in the knowledge graph,lacking the neighborhood information of applied entities.In this thesis,considering the role of neighborhood information on the expression of central entities,the attention mechanism is introduced in the encoder stage.Before aggregating the neighborhood entities,in order to reduce the computational burden of the massive neighbors,the vector representation of the entity and its k neighbors are extracted into the decoder,and the neighborhood information is aggregated to obtain the updated entity representation;finally,the new vector representation is input into the decoder model for link prediction.The NALP algorithm has completed link detection on three datasets: FB15 K-237,WINE and WN18 RR.The experimental results show that the method achieves higher accuracy compared with the baseline model,proving the effectiveness of the method.