Estratégias de Undersampling para Redução de Viés em Classificação de Texto Baseada em Transformers

Automatic Text Classification (ATC) in unbalanced datasets is a common challenge in real-world applications. In this scenario, one (or more) class(es) is overrepresented, which usually causes a bias in the learning process towards these majority classes. This work investigates the effect of undersampling methods, which aim to reduce instances of the majority class, on the effectiveness of recent ATC methods. Through a systematic mapping of the literature, we selected and implemented 15 undersampling strategies. We also propose two new strategies and compare all 17 methods using RoBERTa as sentiment analysis classifier. Our results suggest that a set of undersampling approaches is capable of significantly reducing the learning bias of ATC methods towards the majority class on imbalanced datasets, without incurring any effectiveness loss, and with improvements in efficiency and reduction of carbon emissions.