Análise de sentimentos em tweets utilizando modelos tradicionais, deep learning e RoBERTa

Abstract

This research aims to develop a sentiment analysis framework using different classification models applied to short and informal texts, such as those commonly found on social media platforms like Twitter. The study covers traditional supervised learning algorithms, including LSTM (Long Short-Term Memory), LinearSVC, Logistic Regression, BernoulliNB, and Decision Tree, as well as more advanced models based on the Transformer architecture, with particular emphasis on RoBERTa. These models are evaluated based on their ability to accurately identify positive, negative, or neutral sentiments, considering different preprocessing, vectorization, and textual representation techniques. The proposal seeks to compare these models across various linguistic contexts and structures, highlighting their advantages and limitations within the domain of Natural Language Processing (NLP). Additionally, optimization techniques are employed to fine-tune hyperparameters, aiming to improve model performance on a dataset consisting of one million tweets. The RoBERTa and LSTM models achieved the best accuracy results, with 89.84% and 79.13% respectively, outperforming and remaining competitive with recent results in the literature.