基于历史学习和关系注意力的时序知识图谱推理

Reasoning about Temporal Knowledge Graphs Based on History Learning and Relational Attention

时序知识图谱在传统静态知识图谱上进一步引入了时间维度,由此引出时序知识图谱推理任务,旨在推理未来事件或缺失事实的实体或关系。针对大多数时序知识图谱推理模型存在没有充分利用历史事实和关系关联分析的问题,该文提出一种基于历史学习和关系注意力的时序知识图谱推理方法(简称HLRA)。为了充分利用历史事实,使用多层感知机学习历史事实中的时间戳权重,并结合复制模式的思想,编码出具备时间权重的历史语义偏移向量,在此基础上关联查询信息得到历史学习评分。另一方面,使用自注意力机制分析关系间的关联,将计算出的关系间注意力分数作为影响因子,并将之加权到实体预测中得到关系注意力的得分。最终,通过结合两个分数以获得综合的置信分数。在ICEWS18、ICEWS14、YAGO和GDELT等数据集上的实验结果表明,HLRA模型在MRR、Hits@1、Hits@3和Hits@10等评价指标上较次优模型获得1%至4%的提升,有效提升了时序图谱推理的能力,是一种效果更好的模型。

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 associations,we propose a temporal knowledge graph reasoning method based on historical learning and relational attention(abbreviated as HLRA).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%improvement 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.