基于时间演化图卷积网络的舆情热点内容预测

有效预测舆情事件的热点内容有利于提高对舆论导向的把控能力和对公众诉求的预判能力. 然而，现有的舆情预测工作大多关注事件整体趋势指标或情感极性的演变预测，鲜有针对舆情事件热点内容的预测研究. 为解决以上问题，本文提出一种基于时间演化图卷积网络的舆情热点内容预测方法：以舆情事件的热点词作为预测对象，首先，通过演化图卷积网络学习各时间片词语的空间关联关系；然后，使用门控循环单元捕捉各时间片词语特征的时序变化；最后，通过全连接层进行输出，实现对舆情事件热点词的预测. 以微博上两个不同的舆情突发事件的相关文本作为数据集，与两种现有热点词预测方法开展对比实验. 实验结果表明，该方法在两个数据集上的精确率分别达到51.21%和50.98%，召回率分别达到50.17%和48.15%，F1值分别达到50.68%和49.52%，均高于两种对比方法，能够更好地完成舆情事件中热点词的预测.

A Temporal evolving graph convolutional network for Public opinion prediction in emergencies

Public opinion prediction is one of the key solutions to improve the ability to guide public opinion in emergencies. However, most of the existing public opinion prediction work focuses on the trend indicator or sentiment polarity of events ,while little attention paid to the prediction of hot words and topics in specific events. In this paper, a temporal evolving graph convolutional network for public opinion prediction in emergencies is proposed, in which the hot words associated with specific events are taken as the object of public opinion prediction. Our approach combines evolving graph convolutional network with gated recurrent unit: the former is used to learn the dynamic spatial correlation between words and the latter is used to capture the temporal changes of words, the hot words of an emergency in the next time period is then predicted through full connection layer output. To validate the proposed method, we selected discussion texts related to two emergencies on Weibo as the dataset, and conducted comparative experiments with two existing hot word prediction methods. The results show that our method achieved higher precision, recall, and F1-score in both emergencies, with precision of 51.21% and 50.98%, recall of 50.17% and 48.15%, and F1-scores of 50.68% and 49.52%, respectively. These results demonstrate that our proposed method is effective inpredicting public opinion during emergencies