Análisis de la percepción ciudadana sobre seguridad pública y políticas gubernamentales en ecuador: basado en ciencia de datos, para la toma de decisiones

Abstract

Abstract: A widening gap between perceived social insecurity and official statistics is further eroding trust in institutions in Ecuador. Based on the CRISP-DM approach and the social network X (Twitter), we present an artificial intelligence (AI) system for timely assessment. We propose a hybrid Natural Language Processing (NLP) architecture: lexical algorithms combined with the performance of a Transformer model (Multilingual BERT), tuned with local data. Additionally, an LLM model (GPT-4) was included to generate automated contextualized interpretations. In sentiment classification, the model achieved 98% accuracy and an F1 score. A final solution was also examined, featuring an interactive panel designed to geolocate events and link measures of violence with social perceptions. This work demonstrates that the customization of Deep Learning models is fundamental to transforming unstructured digital discourse into strategic tools for government decision-making.