Abstract: 【Purpose/significance】Education entrance examination is the focus of public attention, which is easy to trigger public opin⁃
ion events.It is very important to monitor and judge the development trend of related network information dissemination, discover po⁃tential public opinions and deal with them in time to ensure the safety of examination and maintain the reputation of universities.【Method/process】By combining BERT and BiLSTM to build a deep neural network model, this study analyzes the emotional polarity of the text of the mainstream media social platform data during the postgraduate re examination.TextRank algorithm is used to extract popular keywords of different emotional polarity generic texts, monitor potential public opinion and put forward management sugges⁃tions.【Result/conclusion】The empirical results show that the model can effectively extract the hot topic information under different emotions, find the potential hidden dangers and public opinion focus, and provide reference and practical basis for universities to moni⁃tor network public opinion.【Innovation/limitation】Compared with the traditional methods, the pre training language model based on
BERT can effectively solve the limitation that the model cannot accurately represent the complex relationship between different sen⁃tences due to the small amount of data. Meanwhile, BERT can model text in two ways, capture the characteristics of the relationship be⁃tween different sentences, and improve the accuracy of text emotional topic mining.