Research On Knowledge Completion Methods Based On Multilayer Attention And Temporal Graph Neural Networks

In recent years,knowledge graphs have been widely used in the field of artificial intelligence.Since the existing knowledge graph data mainly comes from the web and is often constructed by manual or semi-manual means,there are often problems of missing data in the process of data acquisition and construction,which brings challenges to the further application of knowledge graphs.Therefore,how to realize the complementation of knowledge graph data and ensure a certain degree of accuracy has become one of the current research hotspots in the fields of data governance and artificial intelligence.At present,researchers have proposed numerous methods to make use of existing information to complement unknown facts in knowledge graphs,but these methods generally suffer from the problem that they can only learn a single feature representation for entities and relationships,but cannot accurately represent the semantics in the current context.To solve the above problems,this paper provides an in-depth study of knowledge-completion methods.To address the incompleteness of static knowledge graphs,the Bayesian and multilayer attention-based complementary method BRHA is proposed.This method treated the type information and neighborhood information of entities as hierarchical structures,grouped them by relationships,and calculated the attention weights of each type of information within the group independently.Then the type information and neighborhood information of entities are encoded as prior probabilities and the instance information is encoded as likelihood probabilities,and the two are combined according to Bayesian rules.The BRHA method can effectively improve the accuracy of link prediction in static knowledge graphs.To address the incompleteness of the temporal knowledge graph,a complementary method TGNN based on temporal graph neural networks is proposed.This method encoded the specific query substructure of the temporal knowledge graph by using a CNN with an attention mechanism to assign different weights to different types of information,and a GNN with an attention mechanism to focus on the time displacement between each event and the query timestamp,which in turn encoded the specific query substructure of the temporal knowledge graph.The method considered the prediction learning of tail entities from a global entity set and the prediction learning of tail entities from a specific,smaller historical entity set,respectively,and jointly considered both probabilities to obtain the probability of the final predicted tail entities.The TGNN method can effectively improve the accuracy of link prediction in temporal knowledge graphs.For the static knowledge-completion task,the MRR(Mean Reciprocal Rank)metrics in this paper improved 14.4% over Conv E and 10.7% over TKRL in the FB15 k dataset,proving that the BRHA method can effectively improve the accuracy of static knowledge-completion.As for the temporal knowledge completion task,the MRR metrics in this paper in the ICEWS14 dataset improved 2.74% over T(NT)Compl Ex and 6.14% over DE-Simpl E,proving that the TGNN method has good performance in the temporal knowledge graph link prediction task and is interpretable.