Analysis of longitudinal Facebook posts before and after a tornado using large language models

Abstract

The tornado outbreak in the US in 2011 was one of the worst natural disasters in recent history. It left a big mark on communities and sparked a lot of conversation on social media. Based on Facebook posts made during the timeline of the disaster, this thesis presents a new dataset that shows the timeline and posts of public sentiment and behavioural variations during this important event. The main goal of this research is to examine how people's feelings and actions changed during the tornado outbreak as reflected by their Facebook posts. It will give us a full picture of how user opinion changes in real time. Large language models (LLMs) like GPT, BERT and twitter-sentiment-pl are used in the study, along with more advanced methods. I used topic modelling and hybrid mood analysis. These models are used to sort user feedback into groups, find new themes, and keep track of changes in public opinion during tornado timeline (before, during, and after the event). The suggested mixed method uses text features, punctuation cues, and lexicon-based analysis to make sentiment forecasts more accurate and easier to understand. In addition, the study includes a comparison of LLMs, checking how well each model works on the brand new dataset. I fine-tuned the LLMs from their basic model according to my database. So that it can predict more accurate sentiment of the Facebook posts. Using explainable AI (XAI) techniques in this study makes model decisions clear by pointing out important factors that affect classification. Behavioural and emotional changes are shown over time. The study discusses different stages of the public's reaction. The results show that people's feelings and actions changed a lot over time. The sentiments depend on how bad the storm was and how quickly official information spread. By comparing LLMs, we can see what their pros and cons are when it comes to handling disaster-related social media data. This thesis adds a new dataset and a complete framework for studying how people's feelings change during natural disasters. It shows a way to do sentiment analysis in real time so that we can use the technique to prepare a better and quicker disaster management system. This has important effects for emergency response plans and planning for real-time crisis communication.