Authors - Maria Jose Cantos Cedeno, Kevin Michael Mero Ramirez Abstract - The spread of disinformation through social media, messaging apps, and other digital channels is a growing problem in Ecuador, due to the limitations of manual fact-checking processes in the face of the high volume of information. In this context, Transformer-based models are presented as high-potential solutions for detecting fake news across various domains and languages. The objective is to comparatively evaluate Transformer architectures pre-trained using fine-tuning techniques for the automatic classification of fake and real news in the Ecuadorian context. The CRISP-DM methodological framework was applied to guide the development of deep learning models. A balanced dataset of 5,000 news items in Ecuadorian Spanish was constructed, equally distributed between real and fake news. Data processing was carried out through a 12-stage sequential pipeline to reduce noise, prevent data leakage, and preserve relevant semantic features of Spanish. Furthermore, the models were trained using homogeneous hyperparameters and evaluated using various metrics employed in the scientific field. As a result, all models exceeded 89% accuracy; BETO achieved the highest precision, mBERT the best recall, and DistilBERT the highest computational efficiency. It is concluded that Transformer architectures proved to be scalable, effective, and viable solutions for the automatic detection of fake news in Ecuador.
Open Zoom