Reasoning about Temporal Knowledge Graphs Based on History Learning and Relational Attention
Temporal knowledge graph further introduces the time dimension on the traditional static knowledge graph,which leads to the temporal knowledge graph reasoning task,aiming at reasoning about future events or entities or relations with missing facts.Aiming at the problem that most temporal knowledge graph reasoning models have not fully utilized the analysis of historical facts and relational associ-ations,we propose a temporal knowledge graph reasoning method based on historical learning and relational attention(abbreviated as HL-RA).In order to make full use of historical facts,we use multilayer perceptron to learn the timestamp weights in historical facts,and combines the idea of replication patterns to encode historical semantic offset vectors with temporal weights,on the basis of which historical learning scores are obtained by associating query information.On the other hand,we use the self-attention mechanism to analyze the association between relations,use the calculated inter-relationship attention score as an influence factor,and weight it to the entity prediction to get the relationship attention score.Ultimately,the two scores are combined in order to obtain a composite confidence score.Experimental results on ICEWS18,ICEWS14,YAGO and GDELT datasets show that the HLRA model obtains 1%to 4%im-provement over the suboptimal model in evaluation metrics such as MRR,Hits@1,Hits@3,and Hits@10,which effectively improves the ability of temporal graphical inference,and it is a more effective model.
国家科技期刊平台
NSTL
万方数据
