Classificação de produtos com base em descrições textuais

Abstract

Many e-commerce applications have to deal with a large set of product data that needs to be classified into a predefined product category taxonomy. In addi- tion, in some practical scenarios, the data set is volatile, with new products being frequently launched and introduced in these product categories. Product classification has become an essential task for the good functioning of sales platforms in e-commerce environments, facilitating the organization and access to information on the companies’ websites. In this dissertation, we study and discuss efficient and effective methods for product classification. We present a fast and competitive solution for classification based on Language Models to classify products and discuss the use of a classification method proposed in the literature that has been used successfully in other applications, FastText, adapting and studying it in the product classification scenario. We studied ways of combining the proposed methods with product description segmentation, an idea previously used in the literature, and we presented experiments with 3 product databases where we compared the performance of the alternatives studied. The results presented indicate that both the method based on Language Models and FastText present very competitive qualitative results when compared to a classification model based on neural networks that is considered state-of-the-art. The results were obtained with a significant reduction in costs and in the processing time necessary to carry out the experiments in the 3 databases studied.