The use of free text to categorize any type of entity causes, in most cases, difficulties related to the identification of such entities. In the Electronic Fiscal Receipt (“Nota Fiscal Eletrônica”, NF-e), issued for all public purchases in Brazil, products are categorized within the Mercosul Common Nomenclature (NCM). Such an identifier is necessary to calculate taxes, but it is often filled in wrongly, which makes it difficult to detect irregularities in prices and monitor public expenditures. In this context, an automatic product categorization system was developed based on the textual descriptions present in the NF-e. It consists of a categorization tree that follows the NCM product hierarchy, using the Local Classifier per Parent Node pattern. Each node in the tree is trained to encode the textual descriptions in Document Embeddings and then use a supervised classification algorithm to decide the NCM code. Tree nodes are optimized by selecting classification algorithms as well as parameters, testing the performance of various random configurations. In the results, the hierarchical classification presented a higher F1 score than the flat classification experiments and the error propagation problem was mitigated.The use of free text to categorize any type of entity causes, in most cases, difficulties related to the identification of such entities. In the Electronic Fiscal Receipt (“Nota Fiscal Eletrônica”, NF-e), issued for all public purchases in Brazil, products are categorized within the Mercosul Common Nomenclature (NCM). Such an identifier is necessary to calculate taxes, but it is often filled in wrongly, which makes it difficult to detect irregularities in prices and monitor public expenditures. In this context, an automatic product categorization system was developed based on the textual descriptions present in the NF-e. It consists of a categorization tree that follows the NCM product hierarchy, using the Local Classifier per Parent Node pattern. Each node in the tree is trained to encode the textual descriptions in Document Embeddings and then use a supervised classification algorithm to decide the NCM code. Tree nodes are optimized by selecting classification algorithms as well as parameters, testing the performance of various random configurations. In the results, the hierarchical classification presented a higher F1 score than the flat classification experiments and the error propagation problem was mitigated.