Classification of Spanish Fake News about Covid-19 using Text Augmentation and Transformers

Flores A.; Tito-Chura H.; Yana-Mamani V.; Ecos-Espino A. · International Journal of Intelligent Systems and Applications in Engineering (2024)

Resumen

This paper presents the results of five models based on transformers such as DistilBERT, ALBERT, BETO, DistilBETO, and ALBETO for the classification of fake news about covid-19 in the Spanish language. Two text augmentation processes based on GPT-3 are compared, the first TA1 consists of the most common way of increasing the records of the training data, that is, increase all training data; and the second TA2, is more selective in the sense that it only increases the records that could not be learned by the models in the training phase, thus optimizing the training time of the models with respect TA1. The results show that both text augmentation techniques allow improvement, however, TA2 has a better performance in the models based on the Spanish language such as BETO, DistilBETO, and ALBETO, improving on average 1.15%, 11.12%, 2.44%, and 7.50% in terms of accuracy, recall, precision and f1-score respectively.

Instituciones: Universidad Nacional de Moquegua