Análise de sentimentos no domínio do e-commerce utilizando modelos de linguagem

Abstract

Natural Language Processing (NLP) is a subfield of computer science, specifically within Artificial Intelligence, aimed at providing computers with the ability to comprehend words, both in text and speech, similar to how humans do. One domain where NLP has been extensively applied is e-commerce. When consumers make purchases, they often share opinions about products, enabling companies to identify areas for improvement through the emotional content of these reviews. Within the scope of NLP, the task of Sentiment Analysis aims to address this issue by classifying the emotional content behind these opinions. In this context, the primary objective of this work is to compare different approaches of neural networks based on transformer architectures. Notably, the BERT (Bidirectional Encoder Representations from Transformers) model stands out with a remarkable evolution in transformer architecture. BERT, along with its variants, has demonstrated prominence as a class of models that surpass sequential information processing. Among the variants of BERT, BERTimbau, specifically trained for Brazilian Portuguese, stands out as a benchmark model in tasks such as semantic similarity, textual inference, and named entity recognition. This study conducts fine-tuning of the BERT, RoBERTa, and BERTimbau models for sentiment classification, utilizing the B2W-Reviews corpus. The goal is to establish a benchmark in the context of Sentiment Analysis for the Portuguese language.