Maestría en Analítica Aplicada · 2026

Clasificación automática de posts de X para la detección de incidentes durante ciclones tropicales en la República Dominicana

Resumen

This study focuses on classifying X posts to detect incidents during tropical cyclones in the Dominican Republic. To this end, we compile a dataset of 14,321 posts spanning five events between 2017 and 2023: María (2017), Fred (2021), Grace (2021), Fiona (2022), and Franklin (2023), annotated with ten categories: power grid affected, flooding, injured or deceased person, missing person, bridge affected, road affected, housing affected, landslide, other incident, and irrelevant. After preprocessing, BETO (a BERT-based transforme pretrained in Spanish) is trained on the older events, reserving the most recent one for final evaluation to approximate a deployment scenario. In evaluation, the model achieves macro-F1 ˜ 0.69, with macro-precision ˜ 0.68 and macro-recall ˜ 0.77. It performs best on Flooding (F1 ˜ 0.85) and Bridge affected (F1 ˜ 0.81), while showing limitations in underrepresented classes (Landslides, F1 ˜ 0.54) and in the heterogeneous category Other incidents (F1 ˜ 0.45). Overall, the results provide empirical evidence of the potential of automated classification to generate complementary early warnings that support emergency management in the country.