Research On Knowledge Graph Completion Algorithm Based On Graph Neural Network

Knowledge graph has become an indispensable foundation in intelligent scenes such as question answering and text generation due to its rich connection characteristics.The effect of these downstream tasks depends on the completeness of the knowledge graph.In reality,the knowledge graph constructed manually has outstanding defects,so it is vital to complete these incomplete triples.With the rapid update of the real-world information,many unknown entities in the knowledge graph are generated,which is called OOKB(Out Of Knowledge Base).OOKB challenges the existing knowledge graph completion technologies and has become a research difficulty in this field.This thesis analyzes the OOKB problem from the perspectives of relationship prediction tasks and entity prediction tasks.It discusses the feasibility of solving the OOKB problem by relying on the subgraph structure and relationship type information as the structural characteristics of unknown entities.Then,this thesis realizes the above method through Graph Neural Network.The main research work of this thesis is as follows :1.This thesis proposes a knowledge graph completion model based on subgraph structure decoupling.Most existing models directly extract features from subgraph structures without considering semantic differences in different structures.Our model decouples entity features and realizes the aggregation and updating of relevant semantics in different semantic spaces to achieve the decoupling effect of subgraph structure,which allows our model to consider structural features with higher semantic correlation in the prediction.Our model has achieved good results on WN18 RR,FB15k-237,and NELL-995,proving the improved point's effectiveness according to the comparative experiment.2.This thesis proposes a knowledge graph completion model based on relation type optimization.Existing models have the problem of insufficient utilization of relational information.Our model optimizes relational features by graph autoencoders to integrate relational knowledge and improves the aggregation method for generating entity features,which considers the preference of specific prediction tasks to different relationships.Our model has improved Hits@n,MR,and MRR indicators compared with the baseline method.This result shows that incorporating more relational information helps the model generate higher quality entity features.3.Based on this thesis' s knowledge graph completion model,a knowledge graph system for personal relationships is designed and implemented.The system provides relationship prediction,entity prediction,and visual display.The feasibility of the above model to solve the OOKB problem in practical application is verified by the application on the personal relationship dataset.