An Evaluation of Machine Learning Models for Analyzing Disaster-Related Tweets

This paper investigates supervised machine learning approaches for classifying Twitter posts as disaster-related or not to support improved emergency response. A comparative analysis of six models is conducted, including GAP 1d, GRU, RNN, LSTM, BERT, and SBERT architectures, developed using a dataset of over 10,000 tweets. With methodical improvements to generalization, the promise of these natural language processing techniques for extracting accurate, meaningful signals from noisy social media data can better assist disaster management teams, news agencies, community support groups, and other stakeholders when it matters most during crisis events. Rigorous enhancement of reliability remains necessary to transform potential into robust, stable performance amidst volatility.