Linking Opinion Dynamics and Emotional Expression in Online Communities: A Case Study of COVID-19 Vaccination Discourse in Japan

TL;DR

The paper tackles how opinions, emotions, and social communities co-evolve during COVID-19 vaccination discourse in Japan. It integrates macro-level time-series, meso-level community emotion profiles, and micro-level individual emotion shifts by combining SVM-based opinion detection, ensemble-temporal community tracking, and ChatGPT-based emotion classification within a 2020–2022 Twitter dataset spanning ~$40$ million original tweets and $80$ million retweets; the key metric $O_t^i$ is used to label individual leaning. The study reveals phase-specific emotional dynamics and distinct community-level emotion patterns, showing that anti-vaccine shifts are associated with larger increases in Anger and larger decreases in Fatigue, Vigor, and Neutral, while pro-vaccine shifts display different emotional trajectories. These findings have practical implications for public health messaging, suggesting that monitoring and addressing emotional tone—particularly anger and confusion—alongside content could improve vaccination communication and reduce hesitancy.

Abstract

Social media discourse on COVID-19 vaccination provides a valuable context for studying opinion formation, emotional expression, and social influence during a global crisis. While prior studies have examined emotional strategies within communities and the link between emotions and vaccine hesitancy, few have investigated dynamic emotion changes across collective, community, and individual levels. In this study, we address this gap by conducting an integrated analysis of the evolving collective emotions, community affiliations, and individual emotion changes associated with opinion shifts. Our results show that collective emotions exhibit distinct trends in response to vaccination progress. Emotional compositions differ across communities and respond dynamically to changing pandemic circumstances, potentially reflecting the communities' influence on users' opinions. At the individual level, users shifting to pro-vaccine opinions display markedly different emotional changes compared to those shifting toward anti-vaccine opinions. Together, these findings highlight the central role of emotions in shaping users' vaccination opinions.

Figures (4)

• Figure 1: The temporal evolution of collective opinion, social communities, emotions and the hot topics in three vaccination phases. (a) Fraction of users holding pro-, neutral and anti-vaccine opinions b5. The y-axis on the left represents the fraction of users with different opinions, while the y-axis on the right represents the total number of users who are active in each time window. (b) Fraction of users affiliating to six communities. b5 (c) Trend of the fraction of emotions. The black dotted line highlights the boundaries between two vaccination phases.
• Figure 2: Weekly fraction of emotions by communities. Each sub-figure represents one community. The black dotted lines represent the boundaries between two vaccination phases.
• Figure 3: K-means clustering of users who changed opinions. (a) An example of constructing emotion vectors and calculating the increment of emotion for each cluster. (b) Emotion-based clusters for users shifting from neutral or anti-vaccine to pro-vaccine. (b) Emotion-based clusters for users shifting from neutral or pro-vaccine to anti-vaccine. Bar charts show emotion fractions before and after the opinion change; line charts indicate K-means inertia (black curve as represented by the first y-axis) and difference of inertia (grey curve as represented by the second y-axis) used to determine the optimal cluster number via the elbow method. In the line chart, the red dot represents the local minimum point (the turning point), and the red dashed line immediately before the red dot represents the optimal number of clusters that we adopt.
• Figure 4: Detecting peaks from time series. (a) Remove seasonality at a 10-minute interval. The black line shows the original time series which shows strong seasonality with lower activity in midnights. The red line represents the de-seasonalized time series. (b) Up- and down-trend segmentation. The green colors represent upward trends, orange represent downward trends but with weaker magnitude, and red represent downward trends from a strong magnitude.