基于改进XLNet算法的震后舆情分析研究——以甘肃积石山县6.2级和新疆乌什县7.1级地震为例  

Post-earthquake public opinion analysis based on improved XLNet algorithm:a case study of the Jishishan,Gansu M 6.2,and Wushi,Xinjiang M 7.1 earthquakes

作　　者：郑通彦[1] 王尅丰 黄猛[2] 张淞 周文涛 游巧 刘帅[2] ZHENG Tongyan;WANG Kefeng;HUANG Meng;ZHANG Song;ZHOU Wentao;YOU Qiao;LIU Shuai(China Earthquake Networks Center,Beijing 100045,China;Institute of Disaster Prevention,Sanhe 065201,Hebei,China)

机构地区：[1]中国地震台网中心,北京100045 [2]防灾科技学院,河北三河065201

出　　处：《地震工程学报》2024年第4期955-964,共10页China Earthquake Engineering Journal

基　　金：河北省地震灾害防御与风险评价重点实验室开放基金-面上基金项目(FZ223101)。

摘　　要：震后对网络舆情信息的监控与分析,对于相关部门开展震灾应急救援、掌握救灾动态、稳定民众情绪具有重要意义。为解决震后舆情信息数据量大、语言多义性等问题,文章使用自回归模型(XLNet)作为文本向量化表示层,将社交媒体地震数据文本转化为包含上下文语义信息的媒体数据词向量,同时,使用双向门控循环单元(BiGRU)网络作为特征提取层,把词向量序列输入到BiGRU层,提取社交媒体地震数据的文本特征;将初步提取特征的文本输入到注意力机制层(Attention),进一步提取更为重要的情感类别特征,并对重要特征进行权重强化,构建基于网络地震应急处置信息改进的XLNet-BiGRU-Att地震舆情情感分析模型;最终,获得社交媒体地震数据的舆情态势。相比传统的XLNet模型,文章模型在甘肃积石山县6.2级与新疆乌什县7.1级地震的舆情情感分析中能够准确\,快速捕捉长短文本数据特征,分析舆情态势,情感分析准确率分别提升到92.45%和93.42%。The monitoring and analysis of public opinions on online platforms after earthquake aid is highly significant in emergency rescues,understanding disaster relief dynamics,and stabilizing public emotions.However,it is difficult to quickly gather and categorize these opinions given the large volume of post-earthquake public opinion data and polysemy of language.To address these problems,we employed the autoregressive model(XLNet)as a text vectorization layer,which converted the text of earthquake-related data on social media platforms into word vectors containing contextual semantic information.The bidirectional gated recurrent unit(BiGRU)network was used as the feature extraction layer,and the word vector sequence was input into the BiGRU layer to extract text features from these data.These texts were then input into the attention mechanism layer to extract features that are categorized based on sentiments that are highly important.The weights of important features were enhanced to construct an improved XLNet-BiGRU-Att sentiment analysis model based on the information gathered from online earthquake emergency responses.Finally,the public opinion situation of these data was obtained using the model.Compared with the traditional XLNet model,the proposed model yields higher accuracy and can more quickly capture the characteristics of both short and long text data gathered from the public opinion sentiment analysis of the Jishishan and Wushi earthquakes.We successfully increased the sentiment analysis accuracy to 92.45%and 93.42%for the Jishishan and Wushi earthquakes,respectively.

关 键 词：舆情分析 XLNet BiGRU 甘肃积石山 新疆乌什 

分 类 号：P315.9[天文地球—地震学]