Deep Learning Reveals Patterns of Diverse and Changing Sentiments Towards COVID-19 Vaccines Based on 11 Million Tweets
Over 12 billion doses of COVID-19 vaccines have been administered at the time of writing. However, public perceptions of vaccines have been complex. We analyzed COVID-19 vaccine-related tweets to understand the evolving perceptions of COVID-19 vaccines. We finetuned a deep learning classifier using a state-of-the-art model, XLNet, to detect each tweet's sentiment automatically. We employed validated methods to extract the users' race or ethnicity, gender, age, and geographical locations from user profiles. Incorporating multiple data sources, we assessed the sentiment patterns among subpopulations and juxtaposed them against vaccine uptake data to unravel their interactive patterns. 11,211,672 COVID-19 vaccine-related tweets corresponding to 2,203,681 users over two years were analyzed. The finetuned model for sentiment classification yielded an accuracy of 0.92 on testing set. Users from various demographic groups demonstrated distinct patterns in sentiments towards COVID-19 vaccines. User sentiments became more positive over time, upon which we observed subsequent upswing in the population-level vaccine uptake. Surrounding dates where positive sentiments crest, we detected encouraging news or events regarding vaccine development and distribution. Positive sentiments in pregnancy-related tweets demonstrated a delayed pattern compared with trends in general population, with postponed vaccine uptake trends. Distinctive patterns across subpopulations suggest the need of tailored strategies. Global news and events profoundly involved in shaping users' thoughts on social media. Populations with additional concerns, such as pregnancy, demonstrated more substantial hesitancy since lack of timely recommendations. Feature analysis revealed hesitancies of various subpopulations stemmed from clinical trial logics, risks and complications, and urgency of scientific evidence.