Analysis of public opinions on natural disasters on short-video social media based on the Latent Dirichlet Allocation and prompt-BERT-TextCNN models

In the era of new media, social media platforms have become crucial venues for disseminating public event information and forming public opinion, with insufficient analysis of public opinion evolution on short-video-based social media platforms. Taking the major public emergency event ‘2023 Hebei Floods, China’ on TikTok as an example, this study employs the Latent Dirichlet Allocation (LDA) algorithm to extract hot topics of public opinion to reveal the focus of public attention. Based on combining three-layer prompt learning templates and deep learning, a Prompt-BERT-TextCNN sentiment classification model with multi-level feature fusion is used to explore public opinion's hot topics and sentiments at different stages. Empirical evidence demonstrates that the LDA model categorizes the discourse into four primary thematic clusters. Concurrently, the Prompt-BERT-TextCNN model has exhibited reliable performance in sentiment analysis, achieving an accuracy rate of 0.901 and an F1 score of 0.899. These metrics underscore the model's effective capability in sentiment recognition, rendering it a potent tool for discerning the sentiment trend of textual data. The findings provide scientific evidence for supervisory authorities and supporting organizations to implement differentiated and timely public opinion management based on the thematic hotspots and sentiment characteristics at different stages of public emergency events.