<![CDATA[Emergency incident management in Aceh Province is highly dependent on the speed and accuracy of severity triage, yet the growing volume of text-based public reports has outpaced the capacity of manual sorting processes carried out by human operators. Manual triage is slow, inconsistent, and prone to bias, while the unstructured, informal nature of public emergency report text further complicates rapid and reliable severity assessment, creating a risk that critical reports may be delayed or overlooked. This study proposes a text classification method that reconstructs three-tier severity labels (low, medium, high) from the binary categorical fields of the Disaster Response Messages Dataset, a Kaggle-sourced collection of 26,248 real-world disaster-related messages, using a highest-priority-wins rule-based approach adapted from international triage standards. The initial reconstruction reveals a significant class imbalance, with the LOW class dominating at 81.7% of the training data against only 5.8% for HIGH, a condition addressed through Random Oversampling prior to fine-tuning a RoBERTa-base language model through an eight-stage classification pipeline. The proposed method was trained for 5 epochs and evaluated on 2,629 testing samples, achieving an overall accuracy of 87.71% and a weighted F1-score of 87.33%. The LOW class achieved the strongest performance with an F1-score of 93.85%, the MEDIUM class reached 68.60%, while the HIGH class achieved only 36.73% due to its limited sample size and the strict active-request condition applied during label reconstruction. The resulting model was integrated into a Flask API service as part of a web-based emergency reporting system, demonstrating that automated severity classification can feasibly support, though not yet fully replace, manual triage decision-making in a real operational setting.]]>
Copyrights © 2026
