Research on Sentiment Analysis of Online Public Opinion Based on RoBERTa–BiLSTM–Attention Model

Public opinion comments are important for the public to express their emotions and demands. Accordingly, identifying the public emotions contained in comments and taking corresponding countermeasures according to the changes in the emotions are of great theoretical and practical significance to online public opinion management. This study took a public opinion event at a college as an example. Firstly, the microblogs and comment data related to the event were crawled with Python coding, and pre-processing operations such as cleaning, word splitting, and de-noising were carried out; then, the stage of public opinion was divided into phases based on the daily public opinion sound volume, Baidu index, and key time points of the event. Secondly, for sentiment analysis, a supplementary sentiment dictionary of the event was constructed based on the SO-PMI algorithm and merged with the commonly used sentiment dictionary to pre-annotate the sentiment corpus; then, the RoBERTa–BiLSTM–Attention model was constructed to classify the sentiment of microblog comments; after that, four evaluation indexes were selected and ablation experiments were set up to verify the performance of the model. Finally, based on the results of the sentiment classification, we drew public opinion trends and sentiment evolution graphs for analysis. The results showed that the supplementary dictionary constructed based on the SO-PMI algorithm significantly improved the pre-labelling accuracy. The RoBERTa–BiLSTM–Attention model achieved 91.56%, 90.87%, 91.07%, and 91.17% in accuracy, precision, recall, and F1-score, respectively. The situation notification, expert response, regulatory dynamics, and secondary public opinion will trigger significant fluctuations in the volume of public opinion and public sentiment.