基于LDA与BERT-BiLSTM-Attention模型的突发公共卫生事件网络舆情演化分析

Evolution Analysis of Network Public Opinion of Public Health Emergencies Based on LDA and BERT-BiLSTM-Attention Model

[目的/意义]探索突发公共卫生事件网络舆情发展周期中的主题和情感演化历程,研究影响网民情感波动的因素,为网络舆情有效管控提供决策支持。[方法/过程]结合博文数量的时序特征和生命周期理论进行周期划分,利用LDA模型、BERT-BiLSTM-Attention模型构建研究框架,探究不同周期的舆情主题差异及情感演化。[结果/结论]线下病毒变异演化和线上舆情主题与情感演化具有关联性。在新型冠状病毒变异语料库中,BERT-BiLSTM-Attention模型分类准确率为0.8817,F1值为0.8778,其在情感演化分析上具有优越性。构建的“数据采集预处理、舆情周期划分、主题演化和情感演化到获得策略输出”的全过程分析框架对相关部门有效引导网络舆情提供了决策支持和理论支撑,BERT-BiLSTM-Attention模型能更准确地进行情感分类。[局限]数据源单一,面向时间维度上的演化历程未进行时空结合的演化分析。

[Purpose/significance] Explore the theme and emotional evolution process in the development cycle of online public opinion on public health emergency,study the factors affecting the emotional fluctuations of Internet users,and provide decision-making support for effective control of online public opinion.[Method/process] Based on the time series characteristics of blog posts and life cycle theory,the cycle division is carried out,and the LDA model and BERT-BiLSTM-Attention model are used to build a researching framework to explore the differences of public opinion topics and emotional evolution in different cycles.[Result/conclusion] The evolution of offline virus mutation and online public opinion theme both have a relevance to the evolution of emotion.In the novel coronavirus variant corpus,the BERT-BiLSTM-Attention model has a classification accuracy of 0.8817 and an F1 value of 0.8778,which has certain advantages in the analysis of emotional evolution.The whole process analyzing framework of “data collection preprocessing,public opinion cycle division,theme evolution,emotion evolution to strategic support obtaining” constructed can provide decision-making support and theoretical support for relevant departments to effectively guide online public opinion,and BERT-BiLSTM-Attention model can more accurately classify emotions.[Limitations] The data source is single and only faces with time dimension in the evolution process analysis,lacking of space-time integration analysis.