Abstract: Food safety has always been a focal point of societal concern, and the online public opinion risks triggered by this issue have a negative impact on social security and stability. This paper constructs an analytical framework for the evolution of online public opinion on food safety that integrates topic modeling with sentiment feature enhancement. The framework, based on the life cycle theory, combines the temporal characteristics of microblog comment data for periodization. It applies the latent Dirichlet allocation (LDA) model to analyze public opinion topics in different stages. Additionally, the bidirectional encoder representations from transformers-bidirectional long short-term memory-generalized linear model (BERT-BiLSTM-GLM) is used to identify sentiment tendencies, providing an in-depth exploration of the thematic difference and sentiment evolution patterns of online public opinion across various periods, thereby optimizing public opinion response strategies. To validate the effectiveness of the framework, empirical analyses are conducted based on the dataset constructed by ourselves for the “mixed use of tankers for edible and chemical oil” scandal and the ChnSentiCorp dataset. The results show that the F1 scores of the BERT-BiLSTM-GLM model are 98.36% and 97.63% for the two datasets, respectively, demonstrating its superiority in sentiment evolution analysis. Based on this, the LDA model is used for comprehensive analysis of the thematic evolution of online public opinion on food safety, providing strong decision support and a theoretical foundation for relevant government departments to effectively guide online public opinion.

Key words: food safety; online public opinion; evolution of public opinion; deep learning; sentiment analysis