Abstract
In recent months, COVID-19 has become a global pandemic and had a huge impact on the world. People under different conditions have very different attitudes toward the epidemic. Due to the real-time and large-scale nature of social media, we can continuously obtain a massive amount of public opinion information related to the epidemic from social media. In particular, researchers may ask questions such as "how is the public reacting to COVID-19 in China during different stages of the pandemic?", "what factors affect the public opinion orientation in China?", and so on. To answer such questions, we analyze the pandemic related public opinion information on Weibo, China's largest social media platform. Specifically, we have first collected a large amount of COVID-19-related public opinion microblogs. We then use a sentiment classifier to recognize and analyze different groups of users' opinions. In the collected sentiment orientated microblogs, we try to track the public opinion through different stages of the COVID-19 pandemic. Furthermore, we analyze more key factors that might have an impact on the public opinion of COVID-19 (e.g., users in different provinces or users with different education levels). Empirical results show that the public opinions vary along with the key factors of COVID-19. Furthermore, we analyze the public attitudes on different public-concerning topics, such as staying at home and quarantine.
Introduction
During a global crisis, public opinion is not merely a byproduct of events; it is a critical force that shapes social stability and policy effectiveness. However, capturing the "pulse" of a nation in real-time is notoriously difficult. This study addresses the challenge of quantifying public sentiment during the volatile early stages of COVID-19 in China (January to April 2020). By analyzing nearly one million Weibo posts, the researchers track how sentiment shifted as the virus spread, identify which demographic groups were most vocal, and examine how specific events—such as the confirmation of human-to-human transmission—impacted the collective mood.
Methods
The researchers employed a multi-stage pipeline to process social media data:
- Sentiment Classification: The Fasttext framework was used to build a classifier trained on manually labeled data, categorizing posts as positive, negative, or neutral.
-
Sentiment Index: To provide a single metric for "public mood," a Sentiment Index was defined: $$\text{SentimentIndex} = \frac{\text{Positive} - \text{Negative}}{\text{Positive} + \text{Negative}}$$ This index ranges from -1 (purely negative) to 1 (purely positive).
-
Demographic Inference: Classifiers predicted education levels based on user profiles, while metadata provided geographic and age-related insights.
- Statistical Analysis: Regression analysis tracked sentiment trends over time, and Normalized Spearman’s footrule was used to compare regional GDP rankings against sentiment proportions.
Datasets
The study utilized a massive corpus of Sina-Weibo data from two primary sources:
- CCIR 2020 Data Challenge: Used as "seed data" for training the sentiment classifier (covering Jan 1 – Feb 18).
- Custom Crawled Data: COVID-19 related microblogs and over 710,000 user profiles (covering Feb 19 – Apr 15).
- Total Volume: The final dataset included 99,913 sentiment-labeled posts and 900,000 unlabeled posts for broader analysis.
Results and Insights
Regional Activity and the "Wealth-Negativity" Correlation
Analysis revealed that Beijing and Shanghai discussed the pandemic more intensely than even Hubei (the initial epicenter), likely due to higher social media penetration and economic stakes in those hubs.
A striking finding was that regions with higher GDP per capita tended to post more negative and fewer positive microblogs compared to lower GDP regions. Similarly, users with higher educational backgrounds were more likely to express negative sentiments, suggesting that more educated cohorts might be more critical or prone to sharing concerns during a crisis.
The "Human-to-Human" Turning Point
Public sentiment remained relatively stable until January 20, 2020. The official confirmation of human-to-human transmission caused a sharp, immediate drop in the Sentiment Index across almost all regions, marking a significant shift in the national mood.
Demographics and International Relations
- Gender Shifts: In the early stage, female users were significantly more active. As the crisis progressed, the ratio of male-to-female posts equalized, and female users generally became more positive than their male counterparts.
- Age Demographics: The most active age group was 17 to 34, who were also the most likely to express polarized (either very positive or very negative) opinions.
- International Sentiment Gap: There was a massive disparity in how users viewed the Chinese vs. U.S. governments. The Sentiment Index for "China" was 0.69 (highly positive), while the index for the "U.S. government" was a staggering -0.96 (almost entirely negative).
Conclusion
While the Fasttext classifier achieved a precision of 68% (comparable to the VADER sentiment tool), the authors acknowledge limitations such as selection bias—as Weibo users may not represent the entire Chinese population—and the scarcity of explicit demographic data in user profiles. Future work could involve more sophisticated NLP models like BERT to improve classification accuracy and a deeper investigation into the drivers behind the negative correlation between economic status and sentiment.
Summary updated