An Empirical Comparison of Portuguese and Multilingual BERT Models for Auto-Classification of NCM Codes in International Trade

Classification problems are common activities in many different domains and supervised learning algorithms have shown great promise these areas. The classification of goods international trade Brazil represents a real challenge due to the complexity involved assigning correct category codes good, especially considering tax penalties legal implications misclassification. This work focuses on training process classifier based bidirectional encoder representations from transformers (BERT) for with MCN which official system import export products Brazil. In particular, this article presents results using specific Portuguese-language-pretrained BERT model, as well multilingual-pretrained model. Experimental show that Portuguese model had slightly better performance than multilingual achieving an MCC 0.8491, confirms classifiers could be used improve specialists’ goods.