The COVID-19 pandemic, beginning in December 2019, led to widespread implementation of quarantine and lockdown measures globally. These restrictions significantly altered communication patterns, with social media platforms like Weibo becoming crucial channels for information exchange and social interaction in China. The immense volume of data generated on these platforms presents a unique opportunity to analyze public sentiment and its evolution during the crisis. This study focuses on understanding the fine-grained emotional dimensions of online public opinion in China during the first wave of the pandemic. The concept of "public opinion" itself is complex, evolving from its initial usage in 1558 to its current understanding as a collective sentiment shaped by media and increasingly by social media. While previous research has examined topic trends and general emotional tendencies in online discussions surrounding COVID-19, this study aims to provide a more nuanced analysis by focusing on specific emotion categories and their temporal and spatial variations using data from Weibo, a dominant social media platform in China. The research will contribute to a deeper understanding of how public opinion forms and changes during public health crises and will provide valuable insights for effective public opinion management and emergency response strategies.

Existing research on public opinion during the COVID-19 pandemic has primarily focused on topic discovery and overall emotional tendencies, often categorizing sentiment as positive, negative, or neutral. Studies utilizing Latent Dirichlet Allocation (LDA) and other methods have analyzed the temporal and spatial distribution of topics on platforms like Weibo and Twitter. Some researchers have examined the impact of misinformation and its spread, and others have explored the changes in residents' mental states as expressed through social media. However, a detailed analysis of fine-grained emotional dimensions, encompassing specific emotions like happiness, sadness, anger, and fear, has been lacking. This study addresses this gap by utilizing a fine-grained emotion extraction method to delve into the detailed emotional landscape of online public opinion during the initial COVID-19 outbreak in China.

This study employed a combination of techniques to analyze online public opinion related to COVID-19 in China. First, a large dataset comprising over 45 million Weibo posts from December 1, 2019, to April 30, 2020, was collected. The dataset was then preprocessed to clean and prepare the text data for analysis. The core methodological approach involved two key components: LDA topic modeling and text emotion extraction. The Latent Dirichlet Allocation (LDA) topic model was used to identify and categorize the main themes discussed in the Weibo posts. The selection of the optimal number of topics was determined using two evaluation metrics: perplexity and semantic coherence. Perplexity, a measure of model uncertainty, was minimized, while semantic coherence, an indicator of the semantic relatedness of words within a topic, was maximized. The optimal number of topics was selected based on the lowest perplexity and highest semantic coherence values. To analyze the emotional content, a text emotion extraction method was developed, relying on a dictionary of emotional ontology from the Information Retrieval Laboratory of Dalian University of Technology. This dictionary categorizes emotions into seven categories: "like," "happiness," "sadness," "anger," "fear," "disgust," and "surprise." Each Weibo post was then assigned an emotion vector representing the proportion of each of these emotions present in the text. The emotion vectors were normalized to ensure equal weighting of Weibo posts. By combining the LDA topic modeling results with the emotional analysis, this study examined how different emotions were associated with various topics, providing a comprehensive understanding of the emotional landscape of public opinion during the pandemic.

The study revealed significant fluctuations in public emotions throughout the period under investigation. Positive emotions, like "like" and "happiness," saw notable increases during holidays such as Christmas and the Chinese New Year. Conversely, the announcement of human-to-human transmission and the Wuhan lockdown triggered a sharp rise in "fear," particularly in Wuhan and its surrounding areas. While "fear" gradually decreased over the subsequent two months, it remained significantly elevated compared to pre-outbreak levels. A spatial analysis comparing major Chinese cities with their neighboring cities showed a stronger emotional response to COVID-19 in the major cities, particularly concerning "fear." The analysis of Weibo posts related to the Wuhan lockdown revealed initially negative sentiments like "fear" and "sadness" immediately following the announcement. However, the overall sentiment improved in the days that followed, possibly due to the combined effect of the New Year’s celebrations and increasing confidence in government measures. LDA topic modeling of Weibo posts on the day of the Wuhan lockdown identified six key topics: traffic measures, Chinese New Year celebrations, the epidemic itself, daily life, the supply of essential goods, and social relationships. Analysis of emotional distributions across these topics showed that "fear" was strongly associated with the epidemic-related topic, while "like" and "happiness" were associated with topics like daily life and social relationships. The "materials" topic showed high proportions of "fear" and "disgust", reflecting concerns about the availability of essential goods. The "social relationships" topic exhibited a mixed emotional response, with some expressions of positive sentiment towards the government and others expressing criticism.

The findings of this study highlight the significant role of social media in reflecting and shaping public opinion during a major public health crisis. The fine-grained emotional analysis reveals a dynamic interplay of emotions influenced by key events, public health concerns, and socio-cultural factors such as holidays. The stronger emotional response observed in major cities suggests the importance of targeted public opinion management strategies for these areas. The initial negative sentiments following the lockdown announcement highlight the potential for public anxiety and the need for timely, clear, and reassuring communication from authorities. The subsequent improvement in sentiment demonstrates the potential for effective public health measures and communication to alleviate public concerns and maintain confidence. The study's findings demonstrate that integrating topic modeling and fine-grained emotion analysis can provide valuable insights into the complexities of online public opinion during emergencies. This approach can be replicated for similar events in other contexts and can inform the development of more effective public communication strategies.

This study provides a detailed analysis of online public opinion during the early stages of the COVID-19 pandemic in China using Weibo data. A combined approach of LDA topic modeling and fine-grained emotion extraction revealed nuanced insights into the temporal and spatial variations of public sentiment, demonstrating the power of social media data analysis in understanding public responses to crises. The findings highlight the importance of effective communication strategies, particularly during times of uncertainty, and underscore the need for targeted interventions in major urban centers. Future research could improve the text emotion extraction method by incorporating a broader vocabulary encompassing subtle emotional expressions and contextual cues. Further explorations into short text topic modeling and the use of more sophisticated natural language processing techniques could lead to more robust and comprehensive analyses of online public opinion.

The study's reliance on Weibo data means that the findings may not be fully generalizable to other social media platforms or demographics. The emotional ontology dictionary used in the analysis may not capture the full range of human emotions and sentiments expressed in the data. The analysis focuses on a specific period and geographical region, thus limiting the generalizability to other phases of the pandemic or regions with different social and cultural contexts. Further studies with larger datasets and more diverse data sources are needed to corroborate these findings and expand the scope of the research.