Classificação de produtos em Notas Fiscais Eletrônicas usando descrições textuais não estruturadas

Abstract

This work presents an analysis about the feasibility of using machine learning algorithms to product item classification in Electronic Invoices (NFE), based on the textual descriptions of the products to associate them with the corresponding commercial units. The database used in this work was extracted from a NFE database made available by Public Ministry of Paraiba (MPPB). This research has as purpose to contribute to the challenge of defining exactly which product is being described and what quantity was sold in an NFE, and thus help in a possible automated price inspection. To the classification, first was performed a manual labeling of a data sample, classifying them with the matching unit of measurement. This sample was used in the training of two models to perform a product description classification: a traditional word-based solution, adopted as baseline; and a state-of-the-art character-based solution based on recurrent neural networks. We perform a comparison of the performance of these two approaches.