Estudo comparativo de modelos de análise de sentimento em postagens de redes sociais

Abstract

This work presents a comprehensive comparative study of different sentiment analysis models applied to social media posts, using the SemEval dataset (2013-2017) as a benchmark. The research systematically investigates the performance of six distinct architectures: BERT, Ro- BERTA, BERTweet, ELECTRA, DistilBERT, and GPT-4.1-nano, evaluating the impact of pre- processing techniques and regularization strategies. The methodology implements a modular pipeline based on Object-Oriented Programming principles, facilitating reproducibility and ex- tensibility of experiments. Using 50,343 tweets for training and 12,285 for validation, balancing techniques (Random Oversampling), regularization (dropout of 0.435, weight decay of 0.12, la- bel smoothing), and stratified 5-fold cross-validation were applied. The best result was achieved by the BERTweet model with 74.0% accuracy and 73.90% F1-score, surpassing related work and demonstrating the importance of domain specialization, given its pre-training on 850 million tweets. All Transformer-based models showed superior performance compared to traditional ma- chine learning methods, with statistically significant differences validated by the Kruskal-Wallis test (p-value = 0.0002). The study identifies critical patterns such as overfitting after the first epochs and proposes mitigation strategies including adaptive early stopping. Main contributions include an integrated and modular framework following code best practices, the Sentiment- Predictor system with automatic computational resource detection, and empirical evidence that adequate selection of pre-trained models and regularization techniques are determinant for final performance in social media sentiment analysis.