【摘要】 时序知识图谱补全旨在通过利用随时间变化的已知信息来填补时序知识图谱中缺失的实体。当前的研究大多采用图神经网络来捕捉离散时间快照中的时间、实体和关系特征。然而这些方法往往忽视了知识图谱中关键实体的语义信息和不同时刻的图编码信息。为此该文提出了一种双通道图对比学习的时序知识图谱补全方法。首先，该文通过替换语义相似的实体，构建出反事实时序知识图谱；然后，采用两个不同的时序感知编码器对时序知识图谱和反事实时序知识图谱编码；最后，基于时序知识图谱和反事实时序知识图谱的对比，以及时序知识图谱在不同编码器输出的对比，进行双通道联合对比学习实现时序知识图谱补全。在五个数据集上实验结果显示，该模型在各项评估指标上均优于其他九种最先进的模型，其MRR指标最大领先幅度达到1.76%。更多还原

【Abstract】 Temporal knowledge graph completion aims to fill missing entities in temporal knowledge graphs by leveraging evolving temporal information. Existing approaches often overlook the semantic information of critical entities in knowledge graphs and the graph encoding information across different timestamps. To address this, this paper proposes a dual-channel graph contrastive learning method for temporal knowledge graph completion. First, counterfactual temporal knowledge graphs are constructed by replacing semantically similar entities. Then, two distinct temporal-aware encoders are employed to encode both the original and counterfactual temporal knowledge graphs. Finally, dual-channel joint contrastive learning is implemented through comparisons between the original and counterfactual graphs, as well as between different encoder outputs of the original temporal knowledge graph. Experimental results on five datasets demonstrate that the proposed model outperforms nine state-of-the-art models across all evaluation metrics, with a maximum improvement in MRR reaching 1.76%.更多还原