The effect of negative public opinion events on social networks is underestimated. To address the issue of sentiment-based methods not being able to directly achieve early warning of negative online public opinion, this study proposes a sentiment classification and topic extraction-based approach to public opinion topic modeling. Using negative emotional topics as an entry point, this study shifts from investigating negative public opinion events to examining negative public opinion topics, thus facilitating statistical and quantifiable analysis of such events. Additionally, to address the persistent shortcomings of methods for negative public opinion early warning, we construct a novel early warning evaluation metric, which is known as the public opinion topic arithmetic index (POI). This index comprehensively assesses the developmental trends of public opinion topics across three dimensions: explosion index (EI), sentiment index (SI), and dissemination index (DI).

This study employs the ERNIE 3.0 large-scale language model for sentiment classification. The annotated sentiment dataset is further trained and fine-tuned to obtain the required sentiment classifier. It performs sentiment classification on a COVID-19 Weibo emotional dataset, computing various post sentiments. The topic extraction module uses the TF-IDF algorithm to extract topics. Each noun tag is considered a potential topic, whereas each Weibo post is treated as a document. The TF-IDF method captures frequently occurring words by calculating their frequencies and avoiding less important terms that appear in each document. The TF-IDF topic extraction algorithm extracts topics from negative emotional Weibo posts and identifies relevant topics associated with negative public opinion events. Finally, POI is employed for further analysis based on the extracted public opinion topics. Consequently, early warning is achieved by analyzing negative public opinion topics instead of events. Furthermore, POI comprehensively calculates the effect of negative public opinion topics by combining EI, SI, and DI. EI reflects the growth rate of the current number of textual instances related to negative emotional topics compared to the average number in a previous period; SI mainly reflects the public's emotions and sentiments triggered by public opinion topics; and DI mainly represents the scope and speed of dissemination of public opinion topics. Finally, a comprehensive negative emotional topic public opinion index is derived by calculating the EI, SI, and DI of emotional topics and postdata information, and the topics that exceeded the warning threshold are warned.

The experimental results reveal that the proposed early warning model effectively predicts social media public opinion events. Among the top ten negatively perceived topics ranked based on weight, the earliest warning time exceeds the average outbreak day by 161.01 hours, with an average of 2.1 early warnings. Additionally, the earliest warning time exceeds the average peak day by 261.81 hours, with an average of 5.8 early warnings.

We establish a threshold for triggering the arithmetic index of public opinion topics by modeling and calculating the arithmetic index of negative public opinion topics in this study. This enables us to exclude negative topics and corresponding public opinion events that surpass the threshold, thereby achieving early warning for topic-related negative public opinion events. The proposed negative public opinion warning model accomplishes its intended objective by employing sentiment analysis methods for the early detection of online public opinions.