Utilizando técnicas de aprendizagem de máquina e NLP para o problema de Product Matching.

Abstract

E-commerce is a market that grows every year, driven by technological advances that make the purchasing process more convenient and efficient. As a result, the number of sales increases, also increasing the supply of products being sold on the internet. Due to the large volume of offers, the consumer's difficulty in finding a certain product increases, as well as their ability to identify and group similar products, in order to find the best deals. This occurs because, given two identical products, that is, that have the same barcode, they are described in different ways. To this end, there is a technique whose objective is to determine whether two products are equivalent, that is, they correspond to the same entity in the real world, using machine learning techniques, called product matching. In this work, several machine learning models were analyzed, including BERT, in order to choose the best model that will be used to identify products whose description does not match their barcode. The database used will be the product database of invoices issued in the State of Acre, made available by the Court of Auditors of Acre, TCE-AC. At the end of the implementation, the model was able to satisfactorily classify invalid products.