Research On Knowledge Fusion Technology Based On Graph Neural Network

With the continuous advancement of artificial intelligence and natural language processing technologies,Knowledge graph has become an important tool to promote the development of applications such as recommendation system,question answering system and search engine.Various institutions have also launched multiple knowledge graph products,but these knowledge graph products may be built based on different information sources,and the knowledge of certain entities may not be comprehensive.In order to obtain a more complete knowledge graph,knowledge fusion can be performed.The most critical task in this process is to match equivalent entities from different knowledge graphs,which is entity alignment.Current research on entity alignment faces the problem of heterogeneity of entity neighborhoods between different knowledge graphs and the scarcity of seed entity pairs for pre-alignment.In addition,knowledge graphs in the real world are constantly changing,and entity alignment algorithms adapted to dynamic knowledge graphs need to be designed.This paper deeply investigates knowledge fusion technology based on graph neural networks,with the following main work content:1.This thesis proposes a cross-lingual entity alignment algorithm based on neighborhood information in knowledge graphs.To address the problem of entity neighborhood heterogeneity,we design an entity relation representation learning module and a matching module,which utilize the mutually reinforcing properties of relation and entity matching to improve the effectiveness of entity alignment.In addition,we propose a semi-supervised iterative alignment scheme with adaptive thresholds,which incorporates trusted alignment results into the entity seed set to solve the problem of few seed entities for pre-alignment.Multiple experiments are conducted on the public dataset DBP15K,and the results show significant improvements in Hits@1,Hits@10,and MRR scores compared to the baseline model,demonstrating the effectiveness of the proposed improvements.2.This thesis proposes a dynamic entity alignment algorithm based on graph neural networks for knowledge graphs.Existing dynamic alignment models have insufficient utilization of relationship features between entities,and do not take into account the discovery of new relationships and the deletion of outdated relationships during graph changes.Therefore,this paper designs a self-supervised scheme that updates relationship features based on the sets of head and tail entities connected by the relationships.The relationship updates are used to enhance entity alignment during graph changes.Experimental results on public dynamic datasets demonstrate significant improvements in precision,recall,and F1 scores of our proposed model,and it can discover more aligned entity pairs during dynamic alignment.3.Based on the entity alignment algorithm proposed in this thesis,a knowledge graph information management system for the film and television industry is designed and implemented.The system provides users with knowledge information management,knowledge querying,and knowledge visualization functions.We validated the effectiveness of the algorithm studied in this paper in practical application scenarios.