Predicting disasters using social media platforms is a significant but difficult task in the evolving field of NLP. To tackle this, our research presents a robust NLP-driven approach for analyzing social media data to enhance disaster prediction and improve real-time crisis management. This study introduces a hybrid method integrating large-scale social media (SM) analytics with advanced language models (LLM). Initially, SM analytics are applied to a comprehensive dataset from Twitter (now called X), extracting features such as textual and disaster-specific words. Subsequently, a fine-tuned RoBERTa model is utilized within the LLM approach to capture complex language patterns and contextual relationships in tweets, thus improving disaster prediction. Additionally, Llama 2 is used to evaluate the effectiveness of various LLMs. Several performance metrics were used to assess the hybrid approach, including precision, recall, F1-score, and accuracy. The results were notable: the fine-tuned RoBERTa model achieved an F1-score of 0.84, recall of 0.84, accuracy of 0.86, and precision of 0.85. These outcomes demonstrate that the fine-tuned RoBERTa model surpasses state-of-the-art models like XLNet and Llama 2. This research underscores the essential role of social media analytics in understanding and reducing the impact of disasters, broadening the scope of disaster management, and providing valuable insights for real-time disaster response applications in the future.
Mansoor et al. (Wed,) studied this question.