A large amount of information related to earthquakes is released and widely disseminated on the Internet, but there may also be some misleading views, which may generate a large public opinion risk if emergency countermeasures are not quickly formulated. In this paper, multi-user web crawler technology is used to obtain the text data of public opinion after the Sichuan Luding MS 6.8 earthquake, often the crawled text data contain disturbing information, and the data preprocessing work is completed through the steps of word splitting, de-duplication and labeling. After that, the text is transformed into word vectors that can be read directly by computers using the BERT model. The word vectors are put into the LSTM model for training, so as to realize the dynamic monitoring of public opinion sentiment of Sichuan Luding MS 6.8 earthquake. With the dual support of dataset and evaluation indexes, the model of this paper is evaluated and analyzed. Within 72 hours after the Sichuan Luding MS 6.8 earthquake, the information exposure rises to 532647420660 items, which attracts extensive attention from the society, and with the evolution of time, people’s concern about the earthquake public opinion will gradually decline, which comprehensively outlines the change of public opinion sentiment dynamics. In addition, in the intelligent detection of Sichuan Luding earthquake public opinion sentiment dynamics, this paper’s model detects accurately up to 93.42%, much better than the CNN model, which indicates that this paper’s model is able to guide users to the correct public opinion.
