Classificação automática de códigos NCM utilizando o algoritmo Naïve Bayes.

Abstract

This work consists in a study on modeling a text classifier for automatically categorization of product item descriptions into their appropriate NCM codes. The construction of this classifier was performed using the approach known as supervised learning. The training of this classifier was performed using three sets of data, the first one consisting of NCM transactions from Chapter 20 (Beverages, Liquor and Vinegars), issued by the same retail network; the second composed of transaction items also from Chapter 20, but belonging to different stores; and the third set consisting of transaction items belonging to Chapter 90 (Optical, photographic, cinematographic, medical apparatus and instruments, etc.) and emitted by different establishments. Each of these data sets have different difficulties, were previously labeled and extracted from the database of Consumer Tax Notes (NFC) of the Treasury Department of the State of Rio Grande do Sul. In addition to that, the classifier training also required that the documents used as input to the classifier were preprocessed text processing techniques, with the goal of transforming them into a suitable format for model calibration. During the preprocessing step, tokenization, stop word removal, stemming and vector extraction of characteristics were performed using TF-IDF measure. At the end of the process, the performance of the model was evaluated using cross-validation methodology with 10-folds. The results presented by the proposed approach evidences the ability of the model to correctly classify the instances. For the simples and easier data set, an accuracy of 98% was obtained, while for the medium and difficult sets, the accuracy obtained was 90% and 83%, respectively.