基于多头自注意力机制的CNN地震舆情分析模型研究

浙江省地震局,杭州 310063

深度学习;地震舆情;多头自注意力机制;情感分类

CNN Earthquake Public Opioion Analysis Model Based on Multi-Head Self-Attention Mechanism
XU Xiaotong,CHEN Jifeng,LI Dongping,WU LingJie,LI Huanyu,YAO Di

Zhejiang Earthquake Agency , Hangzhou 310063, China

Deep learning;Earthquake public opinion;Multi-head self-attention mechanism;Sentiment classifi⁃cation

DOI: 10.13512/j.hndz.2024.04.03

备注

随着互联网的发展进步,新媒体平台逐步成为普通公众发布和获取地震灾情信息的首选途径,成为地震相关部门迅速了解当前灾情和公众舆论环境的有效渠道之一。运用网络爬虫技术,搜集震后微博用户公开发表的博文与评论构建数据集并进行预处理,为后续分析和建模奠定基础。通过引入多头自注意力机制优化传统CNN模型,构建基于多头自注意力机制的CNN地震舆情分析模型,丰富特征子空间的多样性,并行处理以及捕捉不同级别的特征和信息,增强模型对地震舆情的理解能力。利用模型对2024年3月7日青海玉树州杂多县5.5级地震进行实例应用,对灾后舆情做了可视化展示。通过实验对比,构建的模型加权平均F1达到92.9%、宏平均F1达到92.1%,能够为地震相关部门在震后快速了解灾情情况和公众舆论环境提供辅助支撑。
With the advancement of the internet, new media platforms are progressively becoming the preferred channels for the general public to release and access earthquake disaster information,as well as one of the effective avenues for earthquake-related departments to promptly grasp the current disaster situation and public opinion. This paper employed web crawling technology to collect post-earthquake Weibo posts and comments from users and con⁃structed a dataset that was then subjected to preprocessing,thus laying the foundation for subsequent analysis and modeling. The paper introduced a multi-head self-attention mechanism to optimize the conventional CNN model, thereby developing a CNN earthquake public opinion analysis model based on the multi-head self-attention mecha⁃nism. The paper enriched the diversity of feature subspace,ensured parallel processing,captured different levels of features and information,and enhanced the model’s ability to understand earthquake public opinion. The model was put into practical application and visualization by analyzing the public opinion following the Zadoi M5.5 earth⁃quake, in Yushu Tibetan Autonomous Prefecture, Qinghai Province on March 7th, 2024. Through experimental comparisons,the constructed model achieved a weighted average F1 score of 92.9% and a macro-average F1 score of 92.1%. These results demonstrate that the model can effectively provide auxiliary support for earthquake-related departments to quickly understand the disaster situation and public opinion environment after the earthquake..
·