Evaluation of Approaches Based on the BERT Model for Opinion Mining about the Cachaça Beverage

Opinion mining is a natural language processing task that aims to classify user opinions expressed on e-commerce platforms, social networks and other media. It is an important tool for decision making, monitoring products/services, detecting trends, developing marketing strategies, among others. Much research has been carried out addressing opinions in the English language. The Portuguese language is still very lacking in linguistic resources aimed at training machine learning models. This work contributes to the evaluation of approaches based on the BERT language model for opinion mining in the Portuguese language, in particular, by creating and evaluating a dataset with labeled data in the domain of the beverage called Cachaça. This is a popular drink in Brazil, and of great economic importance. As a result of the experimental evaluation, the approaches based on the BERT model stood out in relation to two baselines, and in a cross-domain evaluation, they achieved values greater than 0.97 in the F1 metric for classification into 2 classes and 0.64 for 3 classes, in the dataset labeled for the cachaça beverage.