摘要
[研究目的]针对目前舆情事件特征构建存在的问题,考虑事件文本差异性,挖掘更加准确合理的事件深层特征,并用于网络舆情反转预测。[研究方法]基于注意力机制和Transformer等深度学习方法构建了一个事件特征提取模型Doc2EV:对于事件描述文本,提出基于等级掩码注意力的事件描述特征提取模型HMA_EV,生成事件描述特征;对于事件评论文本,使用最新的事件评论特征提取模型NL2ER-Transformer,生成事件评论特征;将提取的事件描述特征和事件评论特征进行融合,生成完整的事件特征并用于舆情反转预测。[研究结果/结论]研究表明,基于Doc2EV模型生成的事件特征向量在舆情反转预测任务上取得97.83%的准确率,高于其它基线模型组合。等级掩码注意力机制有利于捕捉不同平台、不同事件文本之间的关联性和重要性,有利于提取到关键性的事件描述特征;不同事件文本特征适用不同特征提取器进行提取,融合后的事件特征更全面准确;Doc2EV模型能够很好地实现舆情反转预测任务。
[Research purpose]In view of the existing problems in the construction of public opinion event features,considering the difference of event text,this paper aims to mine more accurate and reasonable deep features public opinion events and use them for network public opinion reversal prediction.[Research method]Based on deep learning methods such as attention mechanism and Transformer,an event feature extraction model Doc2EV is constructed:For event description text,an event description feature extraction model HMA_EV based on hierarchical mask attention is proposed to generate event description features;For the event review text,the latest event review feature extraction model NL2ER-Transformer is used to generate event review features;The extracted event description features and event comment features are fused to generate complete event features and used for public opinion reversal prediction.[Research result/conclusion]Research shows that the event feature vector generated based on the Doc2EV model achieved an accuracy of 97.83%on the public opinion reversal prediction task,which was higher than other baseline model combinations.The hierarchical mask attention mechanism is conducive to capturing the relevance and importance between different platforms and different event texts,and is conducive to extracting key event description features.Different event text features are extracted by different feature extractors,and the fused event features are more comprehensive.The Doc2EV model can well realize the task of public opinion reversal prediction.
作者
王楠
杜豪
谭舒孺
李海荣
姜家慧
Wang Nan;Du Hao;Tan Shuru;Li Hairong;Jiang Jiahui(College of Management Science and Information Engineering,Jilin University of Finance and Economics,Changchun 130117;Institute of Big Data and Interdisciplinary Science,Jilin University of Finance and Economics,Changchun 130117;School of Information Science and Engineering,Guilin University of Technology,Guilin 541006;School of Information Engineering,Xinjiang Institute of Technology,Akesu 843100)
出处
《情报杂志》
北大核心
2025年第3期107-118,共12页
Journal of Intelligence
基金
国家社会科学基金项目“基于多源大数据事件融合特征预训练的网络舆情预测研究”(编号:22BTQ048)研究成果。
关键词
舆情事件
舆情反转预测
深度学习
等级掩码注意力
事件特征提取
异质平台
Doc2EV
public opinion events
public opinion reversal prediction
deep learning
level mask attention
event feature extraction
heterogeneous platform
Doc2EV
