Research On Knowledge Graph Completion Method Based On Neural Network

Knowledge Graph(KG)is a knowledge representation method based on Semantic Web technology for describing the relationships among entities(e.g.,people,places,events,etc.).However,knowledge graph usually faces the problem of incomplete information,on the one hand,many triads are missing corresponding entities and relationships;on the other hand,most existing knowledge graph complementation models focus on learning knowledge representation for knowledge graphs in closed environments,ignoring their openness,which requires high resource cost to retrain the whole knowledge graph in order to cope with the changes generated by knowledge graphs.As a result,Knowledge Graph Completion(KGC)has become a research area of great interest.In the open world,new entities and relationships are always generated continuously,and most existing knowledge graph completion models are usually based on the assumption of closed environment knowledge graph,i.e.,the knowledge graph will not change,which makes these models often perform poorly when facing the knowledge graph completion in the open environment.Therefore,how to improve the accuracy and efficiency of knowledge graph complementation,and how to continuously extend the knowledge graph with the changes of the external world are current hot issues.This thesis aims to explore in depth how to enhance the completeness of the knowledge graph from three perspectives,in which the main research includes:1)Text GAT,a knowledge graph complementation method based on Graph Attention Networks(GAT)and textual information,is proposed.In this thesis,we propose a knowledge graph complementation method based on graph attention networks and textual information.The method first extracts the feature vectors of entity description text by Bi-LSTM(Birectional Long Short-Term Memory)model,splices them with the entity embeddings in the triad,then trains the joint vectors by graph attention network,aggregates the neighborhood information,and finally achieves the knowledge map complementation task by decoder.The results of Link Prediction experiments in public datasets FB15K-237 and WN18 RR prove that the model combining multi-source information has better representation capability for entities,which can further improve the accuracy and comprehensive performance of the knowledge graph complementation task.2)CText GAT,a knowledge graph complementation method based on graph attention network and additional information,is proposed.The contextual information embedding is obtained,and then the fusion embedding module based on neural network embeds the entity description information and contextual information into entities and relations,and performs link prediction operation to enhance the complementation effect.By adding contextual information and fusing it with structural information and entity description information,the feature representation of entities and relations is enhanced,and the semantic relationships between entities are better distinguished.3)OWEL,an open-world knowledge graph complementation method based on LSTM(Long Short-Term Memory)and attention mechanism,is proposed.real-life knowledge graphs are constantly evolving,and an open-world knowledge graph complementation method based on LSTM and attention mechanism is proposed for the task of open-world knowledge graph complementation.By improving the OWE model,modularizing it and proposing an enhanced extension method,introducing LSTM and attention mechanism in aggregating word vectors,allowing word sequence elements to be assigned different weights according to the degree of association with different relations,improving the aggregation function,introducing alignment function to align two vector spaces,and thus improving the feature representation of external entities and improving the accuracy of knowledge graph complementation.