Abstract: 【Purpose/significance】To explore an effective deep learning-based sentiment classification method for Weibo texts, and to
study the relationship between the sentiment types of users' reposted speeches and the dissemination of private information about hot
events on Weibo.【Method/process】Four deep learning classification models, BERT, BERT+CNN, BERT+RNN and ERNIE, were se?
lected to set up comparison experiments to validate the models with better performance based on reconstructing a seven-category senti?ment classification corpus. Four hot cases in Weibo are selected to make an empirical analysis from four aspects: sentiment distribu? tion, sentiment word frequency, reposting time and reposting number.【Result/conclusion】Through empirical research, it was found that users' dissemination of private information was rapid and short-lived, and negative emotions such as "Anger" and "Disgust" were dominant in the dissemination. The type of sentiment and expressions differ depending on the subject of the private information.【Inno? vation/limitation】The sentimental characteristics and the connection between spreading private information are studied, which pro? vides a valuable theoretical and practical basis for protecting the privacy information security of social network users. But, the amount of corpus data constructed is not enough for training a high-accuracy deep learning model, and the model is not effective in recogniz? ing texts such as inversion and irony.