摘要
为了充分利用时序知识图谱数据的时间维度特征,解决时序知识图谱时间信息挖掘不充分的问题,提出了一种融合时间编码和时间分布注意力的推理模型TETDM,该模型结合长期历史信息和短期历史信息来捕捉不同模式的历史事件对推理目标的影响。首先,模型将较长一段历史信息中对应实体或关系出现的频率作为推理模型的长期约束条件。其次,模型通过时间感知编码和时间分布注意力机制对短期历史信息进行编码,深入挖掘时序知识图谱数据的时间维度特征。其中,时间感知编码器能够精准捕捉时间信息对实体和关系的影响,从而生成带有时间特征的实体和关系表示。时间分布注意力机制通过在不同子图中建模每个重复事实的注意力,而不是仅仅学习它们的表示,来学习历史重复事件的变量分布。通过编码时间特征对实体和关系的影响以及编码不同时间戳下历史信息的贡献程度,可以获得在时间维度上更为精准的实体和关系的嵌入表示,从而提高模型的推理能力。在ICEWS14、ICEWS05-15、ICEWS18以及GDELT数据集上的结果进一步证实,TETDM提出的时间编码模块和时间分布注意力模块均可以提升模型的推理性能。
To fully leverage the temporal characteristics of temporal knowledge graph(TKG)data and address the issue of insufficient utilization of temporal information,we propose a reasoning model named TETDM(Temporal Encoding and Temporal Distribution Attention-based Reasoning Model),which integrates long-term and short-term historical information to capture the impact of different historical event patterns on inference targets.Specifically,the model incorporates the frequency of entities or relations over an extended historical period as a long-term constraint for inference.Furthermore,it encodes short-term historical information using a temporal-aware encoder and a temporal distribution attention mechanism to deeply explore the temporal characteristics of TKG data.The temporal-aware encoder accurately captures the influence of temporal information on entities and relations,generating temporal feature-enriched representations of entities and relations.The temporal distribution attention mechanism,instead of merely learning representations of repeated facts,models the attention of each repeated fact within different subgraphs to learn the variable distribution of historical repetitive events.By encoding the influence of temporal features on entities and relations,as well as the contribution of historical information across different timestamps,the model produces more precise temporal embeddings of entities and relations,thereby enhancing its reasoning capability.Experimental results on ICEWS14,ICEWS05-15,ICEWS18,and GDELT datasets further demonstrate that both the temporal encoding module[JP2]and the temporal distribution attention module proposed by TETDM can improve the reasoning performance of the model.
作者
董文永
梁智学
周孟强
贾亚洁
DONG Wen-yong;LIANG Zhi-xue;ZHOU Meng-qiang;JIA Ya-jie(School of Information Network Security,Xinjiang University of Political Science and Law,Tumushuke 843900,China;School of Computer Science,Wuhan University,Wuhan 430072,China;School of Cyber Science and Engineering,Wuhan University,Wuhan 430072,China)
出处
《计算机技术与发展》
2025年第6期182-188,共7页
Computer Technology and Development
基金
国家自然基金面上项目(61672024)
国家重点专项研发计划(2018YFB2100500)。
关键词
时序知识图谱
图卷积神经网络
注意力机制
时间编码
知识推理
temporal knowledge graphs
graph convolutional neural networks
attention mechanism
time encoding
knowledge reasoning
