AI-based social media analytics for disaster resilience: A unified framework for reliable information extraction

초록

Social media platforms have become crucial information channels during disasters, allowing affected populations to report damages, request aid, and share real-time updates. While these data are valuable for emergency response, their usefulness is often diminished by irrelevant, biased, or misleading content. This study presents a unified, multi-phase framework that leverages advanced natural language processing to extract credible, actionable knowledge from disaster-related social media streams. The framework begins with relevance detection, utilizing a fine-tuned RoBERTa classifier to filter out non-relevant posts. The refined dataset is then processed through an LLaMA3-based Retrieval-Augmented Generation (RAG) module that performs evidence-grounded verification to identify and exclude biased or misinformation-laden content. Finally, unbiased posts are analyzed using Latent Dirichlet Allocation (LDA) and BERTopic to uncover key public needs such as evacuation, medical aid, and infrastructure damage. Experimental evaluation achieved over 92.2% accuracy in credibility classification and generated coherent, interpretable topic clusters. By integrating semantic filtering, factual validation, and topic modeling into a single workflow, this study provides an interpretable, data-driven approach to enhance disaster situational awareness and support evidence-based decision-making toward a more resilient society.

키워드