TL;DR
This study dives into how discussions about HPV vaccines differ across text, audio, and video social media platforms in China. By analyzing millions of posts, the researchers found that each platform uniquely shapes public perception, with text-based platforms amplifying barriers, audio fostering balanced discussions, and video driving rapid emotional shifts. This work offers crucial insights for public health communicators to tailor their strategies for better vaccine uptake.
What Problem Does This Paper Address?
Despite the proven effectiveness of HPV vaccines, uptake remains a challenge in many regions, including China. Understanding public sentiment and beliefs is crucial for improving vaccination rates. However, social media, a primary source of health information, is not monolithic. Different platforms—text-based (like Weibo), audio-based (like Ximalaya), and video-based (like Douyin)—offer distinct ways for users to engage with information, potentially influencing their perceptions of vaccine susceptibility, severity, benefits, and barriers. Previous research has largely focused on text-based platforms, neglecting how audio and video might uniquely shape health discourse and emotional responses. This paper aims to bridge that gap by conducting a cross-modality analysis of HPV vaccine discussions.
Methods
The researchers employed a multi-faceted approach:
- Cross-Modality Data Integration: Audio and video content from Ximalaya and Douyin were transcribed into text to enable a unified content analysis across all three platforms.
- Health Belief Model (HBM) Framework: Discussions were systematically coded based on the four HBM dimensions: perceived susceptibility, severity, benefits, and barriers. This provided a structured lens to compare how these constructs are discussed across modalities.
- Topic Modeling (BERTopic): To identify representative themes within each HBM dimension and platform, the BERTopic model was used. This involved dimensionality reduction and clustering techniques to uncover hidden themes in the textual data.
- Emotion Analysis: A two-step sentiment detection process was used, first classifying neutral emotions and then categorizing non-neutral posts into positive or negative. This allowed for an analysis of emotional fluctuations over time and across modalities.
- Trend Analysis: Time-series data of HBM dimensions and emotions were analyzed to track temporal variations and identify how events like policy announcements or social incidents influenced discourse on each platform.
Datasets
The study utilized data from three major Chinese social media platforms spanning July 2018 to March 2023:
- Weibo (Text-based): 273,357 posts.
- Ximalaya (Audio-based): 1,228 podcasts.
- Douyin (Video-based): 1,225 videos.
The data was preprocessed to ensure comparability, with audio and video transcribed into text. For Weibo, a stratified random sample was used for cleaning, while Ximalaya and Douyin datasets were manually reviewed due to their smaller size.
Key Insights and Ideas
- Modality Matters for HBM Dimensions: Text-based platforms heavily emphasize "barriers" (like cost and side effects), while audio and video offer a more balanced discussion across all four HBM dimensions. This suggests that different platforms naturally lend themselves to discussing different aspects of health decisions.
- Emotional Resonance Varies by Platform: Video and audio platforms tend to evoke stronger emotional responses compared to text. This is likely due to their more immersive and narrative-driven nature.
- Distinct Topic Focus per Modality: Weibo discussions often revolve around public health events and policy concerns. Ximalaya delves into economic and industrial aspects, while Douyin focuses on personal anecdotes and everyday experiences.
- Temporal Dynamics of Discourse: Major events like policy changes or supply shortages trigger different response patterns across platforms. Weibo shows rapid, short-lived spikes in sentiment, Ximalaya fosters longer, more sustained debates, and Douyin exhibits fragmented, scattered bursts of attention.
Key Results
- Barriers Dominance: Barriers constituted 87.15% of posts on Weibo, significantly higher than susceptibility (4.58%), severity (11.84%), and benefits (12.23%). Ximalaya and Douyin showed more balanced distributions, with barriers still being the most frequent but less overwhelmingly so.
- Emotional Landscape: Weibo had a balanced split between neutral and emotional posts, with negative emotions dominating the latter. Ximalaya and Douyin also showed a significant presence of negative emotions, often linked to susceptibility and severity discussions.
- Case Study Insights: During the nine-valent HPV vaccine coverage expansion, Weibo saw immediate negative sentiment due to cost and appointment difficulties, which quickly faded. Ximalaya discussions lasted longer, with a mix of optimism and frustration. Douyin showed fragmented spikes, with attention shifting to domestic vaccines.
Limitations and Future Work
- Platform Specificity: The study focused on three specific platforms. Future work could explore other platforms within the same modalities (e.g., Bilibili for video, Xiaoyuzhou for podcasts) to assess the generalizability of findings and disentangle modality effects from platform-specific attributes.
- Loss of Multimodal Nuance: Transcribing audio and video into text may have led to the loss of crucial contextual features like tone and emphasis. Future research could benefit from multimodal analysis that integrates text, audio, and visual cues.
- Cultural and Temporal Scope: The study was limited to Chinese platforms and a specific time frame. Cross-cultural comparisons and analyses during health crises could offer deeper insights.
- Advanced LLMs and Multimodal Fusion: Future work could leverage advanced language models (LLMs) for more nuanced emotion detection and employ multimodal fusion techniques to integrate features from different media types for more robust analysis.
- Integration with Other Theories: Combining the HBM framework with other theories like the Theory of Planned Behavior (TPB) could provide a more comprehensive understanding of vaccine discussions.
- User-Level and Network Analysis: Incorporating user-level data and network analysis could reveal how information spreads and identify key influencers.
Summary updated