Fine-Grained Semantic Classification of Disaster-Related Social Media Text for Emergency Management

Disaster-related social media posts often report casualties, rescue needs, infrastructure damage, shelter demand, and local situation changes earlier than formal channels, yet their brevity and noise make operational classification difficult. This study examines whether a practical and reproducible classification pipeline can support fine-grained, emergency-oriented semantic recognition under a deliberately conservative evaluation setting. We convert 14,392 English tweets from CrisisSense-LLM into six actionable semantic categories, partition the data by proxy event groups, and compare a TF-IDF logistic-regression baseline, supervised BERT-base fine-tuning, and zero-shot natural language inference. The evaluation is further extended to mapped HumAID data, a manually reviewed 177-post Chinese boundary-test set, a Chinese-to-English translation bridge, and a fixed-budget selective adjudication simulation. BERT-base achieves the best held-out main-test performance (Macro-F1 = 0.8824), outperforming TF-IDF (0.6133) and zero-shot inference (0.3581), and reaches 0.8132 Macro-F1 on HumAID without retraining. Direct English-to-Chinese transfer is ineffective, whereas multilingual BERT and translation bridging improve Chinese Macro-F1 to 0.2684 and 0.3603, respectively. With 600 reviewed posts, selective adjudication reaches 0.7792 Macro-F1 on the main test set and 0.7153 on HumAID. These findings indicate that the central contribution is not a new model architecture, but an empirically validated workflow that combines supervised fine-tuning, leakage-aware evaluation, external validation, cross-lingual stress testing, and information-driven human review. The novelty therefore lies in the reproducible integration of data mapping, group-aware evaluation, external and cross-lingual stress testing, and selective human review into a single emergency-oriented assessment workflow.