Uso de entidades nomeadas para reconhecimento de marcas em descrições de produtos

Abstract

Price comparators are softwares that provide information about products and prices, enabling searches and comparisons, in order to ease the decision of the best place to shop, acting as helpers in this process. One of those comparators is Ludiiprice, an application for mobile devices that allows search, comparison and monitoring of product prices, that has its database populated through users’ collaboration, in which the users provide the data from the invoices QR Code reading and through the description of products extracted from the internet. This paper’s goal is to train a named entity recognition model to extract the brands on the products’ description, with the goal of increasing the product identification confiability in LudiiPrice, improving the recommendation and comparison processes between products. To achieve this, we built a dataset of 16112 products, extracted from three different e-commerces. It was used to train and evaluate a machine learning model, designed to extract the brands of product descriptions, with which we obtained the f1-score of 91% for the recognition of the beginning of the brands, 77% for the middle of the brands and 82% for the end of the brands. Furthermore, we have developed a new product suggestion system that uses the named entities in addition to cosine similarity, obtaining a hit hate 94%, which is 17.1% higher than that of the previous system, which used only cosine similarity.