【摘要】 传统的时态知识图谱补全方法将事实限定在时间不变的假设下，难以区分具有相似语义的实体，且无法捕捉事实在时间上的演化规律。论文提出一种基于事实演化增强的时态知识图谱补全方法 FeeNet(Fact Evolution Enhancement Networks)，该方法使用时间序列预测模型LSTM学习具有时间感知的实体间的关系表示，并结合经典的TransE方法设计得分函数对事实的合理性进行评估，使用基于复制模式的方法识别重复事实，计算出历史重复性得分，最后综合二者给出对于事实成立与否的最终判断。在含有时态信息的ICEWS14等数据集上的实验结果表明，论文提出的FeeNet方法能够显著提升知识图谱补全和预测任务的性能表现。更多还原

【Abstract】 Traditional temporal knowledge graph completion methods limit facts to the assumption that time is constant,which makes it difficult to distinguish entities with similar semantics and capture the evolution of facts in time. This paper proposes a method FeeNet(Fact Evolution Enhancement Networks)to complete the temporal knowledge graph with enhanced fact evolution. This method uses the time series prediction model LSTM to learn the time-aware representations of relationships between entities,and combines TransE to design a score function to evaluate the rationality of the facts,uses the method based on the replication pattern to identify repeated facts,and calculates the historical repeatability score,the final judgment of whether the facts are true or not is given by combining the two. The experimental results on ICEWS14 and other datasets with temporal information show that the FeeNet model proposed in this paper can significantly improve the performance of knowledge graph completion and prediction tasks.更多还原