A data-driven risk analysis and early-warning model for residential fires based on accident investigation reports

• Proposed SHARP framework predicts residential fire severity from text reports. • LLM-based semantic chunking improves feature extraction from accident reports. • Random Forest model balances 80.4% predictive accuracy with interpretability. • Escalation factors like blocked routes drive fire severity more than ignition. Fire accidents pose a significant threat to public safety. Although official investigation reports are a rich repository of causal information, their unstructured textual format presents a significant challenge to large-scale, quantitative risk analysis. This study aims to construct a Semantic Hybrid Accident Risk Prediction framework (SHARP) by enhancing text data preprocessing techniques for improved risk characterization. The framework maps extracted semantic risk factors to accident severity categories and its effectiveness is validated by evaluating the performance of a Random Forest model across three experimental scenarios. Experimental results demonstrate the framework’s efficacy, with its Full Feature Model (FFM) achieving a prediction accuracy of 80.43%. The results further indicate that ’escalation factors’ that impede escape, such as ’Blocked Evacuation Routes/Exits’, show the strongest statistical association with accident severity compared to initial ’ignition factors’. These findings have significant practical implications, offering a data-driven basis for shifting fire safety management from a reactive posture to a proactive, risk-based prevention strategy.