TL;DR
This paper introduces BertFDA, a novel framework that combines BERT, LDA, and Functional Data Analysis (FDA) to analyze public opinion on COVID-19 vaccines using Weibo data. It reveals distinct concerns between vaccine-supportive and hesitant groups and accurately predicts sentiment evolution, offering valuable insights for public health communication.
What Problem Does This Paper Address?
The COVID-19 pandemic has highlighted the critical need to understand public sentiment regarding vaccines. While global efforts exist to track affective states, there's a scarcity of research that deeply analyzes the topics of discussion and the evolution of sentiment specifically for COVID-19 vaccines, especially differentiating between pro-vaccine and vaccine-hesitant populations. Existing methods often overlook the rich semantic context in social media text and the dynamic, continuous nature of public opinion. This gap makes it challenging for governments and public health officials to craft effective communication strategies and interventions to boost vaccination rates.
Methods
The authors developed a framework called BertFDA to address these challenges:
- Dynamic Topic Clustering (LDA): They used Latent Dirichlet Allocation (LDA) to identify 118 distinct topics from over 2.2 million Weibo posts, capturing the evolving themes of discussion over two years.
- Improved BERT for Sentiment Classification: To better understand the nuances of public opinion, they enhanced a BERT model. This involved an improved pre-training task and the inclusion of a more comprehensive emotional corpus (including Sina Weibo and Baidu Tieba data) to enhance its ability to classify sentiment in complex social media text.
- Functional Data Analysis (FDA): Recognizing that public sentiment isn't a series of discrete points but a continuous evolution, they employed FDA. This technique models sentiment as a smooth, continuous function, allowing for a more accurate representation of dynamic changes and better prediction of future trends. They specifically used Bernstein basis functions for fitting these curves.
- Machine Learning for Prediction: Various machine learning models (AR, SVR, RF, GBDT, Adaboost) were used to predict future sentiment trends based on historical data.
Datasets
- Weibo Data: The primary dataset consists of 2,597,823 microblog posts collected from Sina Weibo between January 23, 2020, and January 23, 2022. After cleaning and preprocessing, 2,353,435 valid posts were used.
- Baidu Index: This was used to gauge public attention and compare it with Weibo post volume.
- Emotional Corpus (TB and TW): For improving the BERT model, they utilized the original Chinese Wikipedia corpus, along with public Sina Weibo and Baidu Tieba emotional corpora (referred to as TB and TW). They also added public health emergency-related emotional annotation sets for in-depth pre-training.
Key Insights and Ideas
- Distinct Concerns: The study revealed a clear divergence in concerns between vaccine-supportive and vaccine-hesitant groups. Supporters focused on vaccine effectiveness and news reporting, indicating group cohesion. Hesitant individuals, however, were more preoccupied with COVID-19 variants and vaccine side effects.
- Sentiment Evolution: Public sentiment towards COVID-19 vaccines was generally positive but showed fluctuations. Negative sentiment declined in the early stages of vaccination but later saw an increase, possibly due to concerns about variants and side effects, before improving as vaccination rates rose.
- FDA for Dynamic Understanding: The application of FDA provided a more nuanced view of sentiment evolution, moving beyond discrete time-series analysis to capture the continuous flow of public opinion. This functional approach is key to understanding subtle shifts.
- Weibo as a Rich Data Source: The sheer volume and emotional content of Weibo posts make it an invaluable resource for understanding public discourse on critical health issues like vaccination.
Key Results
- Topic Modeling: Identified 118 topics from over 2.2 million Weibo posts over 24 months.
- Sentiment Distribution: Out of 2,353,435 analyzed posts, 83.3% were positive and 16.7% were negative.
- Prediction Accuracy: The SVR sentiment prediction model achieved the best performance, with MAE of 0.0294, MSE of 0.0014, RMSE of 0.0376, and MdE of 0.0243. The overall prediction accuracy for vaccine public sentiment was reported to have an average relative error of less than 5%.
- Key Topics: Identified "Global development trend of the COVID-19" and "The progress of China's COVID-19 vaccine research and development" as highly continuous topics.
Limitations and Future Work
- Representativeness of Weibo: While Weibo is popular, it represents only a portion of China's population and may not be fully representative of all demographics, particularly older adults or those in less developed regions. This introduces a potential "digital divide."
- Data Completeness: The study acknowledges limitations in obtaining all relevant microblog posts, and user activity patterns (e.g., peak hours) might influence sentiment representation.
- Platform Administration: The unique internet environment in China could introduce biases in the data and discussion quality.
- Future Work: The authors suggest incorporating geotags for more granular regional analysis, investigating finer time scales, and extending the BertFDA framework to other public health crises. They also emphasize the importance of continuing to track sentiment as COVID-19 becomes endemic and policies evolve.
Summary updated