Presentation description

Effective public communication during a health crisis can save lives, but not all messages are received equally. During the COVID-19 pandemic, government agencies and politicians used social media to communicate risks and recommend protective actions. However, public responses varied: some replies were logical and fact-based ("discussion"), while others were emotional or accusatory ("reaction"). Understanding these differences is key to improving crisis communication.

This project investigates how different types of messages-categorized by topic (risk vs. protective action) and source (government vs. political figures)-influence the type of public response they receive. I began by labeling 2,000 tweet replies as either "reaction" or "discussion," then trained a Gaussian Process classifier that achieved 62% accuracy in predicting response type across 1.27 million tweets. To extend beyond traditional models, I incorporated linguistic features such as mentions of statistics, money, and LIWC categories as proxies for reasoning style.

Currently, I'm exploring the use of LLAMA, a large language model (LLM), to improve classification accuracy. Unlike simpler models, LLAMA can evaluate tweets in context, offering richer insights into tone and intent. This stage involves prompting LLAMA to classify responses and comparing its outputs to both human annotations and previous machine learning predictions. I also analyze how these engagement patterns shift depending on the source (e.g., Democrat vs. Republican) and message type (e.g., primary risk vs. secondary protective action).

Early results suggest that politicians' tweets elicit more emotional and money-related reactions, while government agencies prompt more factual discussions. Additionally, tweets about risk are more likely to generate replies containing statistics than those about protective actions.

By combining manual labeling, traditional models, and large language models, this research sheds light on how people engage with crisis communication and how message design might be improved for future emergencies.