Análise Automática de Satisfação do Consumidor na Plataforma Consumidor.gov

This paper studies the effectiveness of Sentiment Analysis, via the BERTimbau language model, for a consumer complaints dataset from the Consumidor.gov platform. Transfer Learning is employed for the finetuning phase and a manually-labelled test set is designed to evaluate the model. Classification results show a F1-Score of 74% for the "Satisfied" class and 81% for the "Unsatisfied" class.