A prompt-based sentence-embedding and gradient boosting framework for construction accident severity assessment in megaproject safety management

This study develops and evaluates a framework integrating prompt-based sentence embeddings with gradient boosting for construction accident severity assessment in megaprojects. The dataset comprises 5224 structured accident records across six highly imbalanced severity levels. Coded accident attributes were converted into controlled natural-language descriptions and encoded using four pretrained sentence-embedding models. The resulting embeddings were reduced using fold-fitted principal component analysis and combined with the original structured predictors. Four gradient boosting classifiers were evaluated across 16 hybrid configurations using a fixed stratified hold-out test set, three-fold cross-validation on the training data, fold-specific SMOTE, and five random seeds. LightGBM with BGE achieved the highest observed mean hold-out Macro-F1 of 0.3489 and a mean accuracy of 0.6304. The hybrid representation produced a slightly higher mean Macro-F1 than structured predictors alone, at 0.3489 versus 0.3446, while embeddings alone achieved 0.3129. Balanced Random Forest attained a numerically higher mean Macro-F1 of 0.3572 but a substantially lower mean accuracy of 0.4961, revealing a trade-off between class-balanced recognition and overall predictive accuracy. These findings suggest that prompt-based sentence embeddings offer a modest, model-dependent complement to structured accident attributes rather than consistently superior performance. The framework may support preliminary record screening and human-assisted safety review.