Accuracy of Natural Language Processing Techniques in Extracting Fire Alert Information from Social Media Posts (Twitter/X) and Early Warning Time Advantage Compared to Official Reporting Systems

Early detection of natural disasters is critically important for minimizing losses. Social media platforms, particularly Twitter/X, enable users to share real-time observations, thereby constituting an alternative information source to traditional reporting mechanisms. This study aims to evaluate the accuracy of natural language processing (NLP) techniques in extracting fire alert information from Twitter/X posts and to measure the early warning time advantage compared to official reporting systems. BERT, RoBERTa, and BiLSTM-CRF-based classification models were employed; over 185,000 tweets related to forest fires occurring between 2020 and 2025 in Turkey and worldwide were collected and annotated. Preprocessing stages included noise removal, tokenization, named entity recognition (NER), and sentiment analysis. Findings indicate that the fine-tuned BERTurk model classified fire-related tweets with an F1 score of 91.4%. Location extraction accuracy was measured at 84.7%. Temporal analysis results demonstrate that the social media-based detection system provided an average early warning of 22.6 minutes compared to official reporting channels. The study reveals the potential benefits of integrating NLP-powered social media monitoring systems into disaster management infrastructure and offers concrete recommendations for decision support systems.