Use of Bi-LSTM for Emotion Detection via Text Messaging in the Latin American Context

Resumen

Emotional violence in digital environments poses a growing risk in Latin American contexts, where verbal expressions vary significantly by region. This study presents an emotion-detection model based on a Bidirectional Long Short-Term Memory (Bi-LSTM) network, trained specifically on messages collected from Twitter using web-scraping techniques. Unlike prior approaches, the model does not rely on surveys or corpora structured in neutral Spanish; instead, it draws on real, everyday language with contextual emotional content. The methodological pipeline spans from data collection and cleaning to the binary classification of violent versus nonviolent messages. The model achieved an accuracy close to nine out of ten correct predictions in identifying violent messages, with recall similarly high for both classes and an F1 score near 0.9. In addition, the model exhibited semantic capabilities to interpret ambiguous phrases, local expressions, and neutral messages, reducing false positives.