Classificação uni e multirrótulo de sentimentos em textos curtos

Abstract

This work investigates the effectiveness of different classification models in the task of identifying emotions in short texts extracted from social media, adopting both single-label and multi-label approaches. To this end, a corpus was constructed comprising 2,209 real comments (natural dataset) and 1,669 samples artificially generated by LLMs (synthetic dataset), aiming to mitigate class imbalance. Traditional models such as Support Vector Machine (SVM), Multilayer Perceptron, BERT, and a Large Language Model (ChatGPT) were evaluated, considering both single-label and multi-label classifications. The samples underwent lexical preprocessing and tokenization, and the models were trained with different data compositions. The results indicate that the use of synthetic data and preprocessing techniques significantly improved model accuracy and generalization, especially for minority classes. The BERT model achieved the best performance in the single-label approach (accuracy of 73.09%), while the LLM obtained the best performance in the multi-label classification (Jaccard Index of 69.91%). The findings highlight the potential of combining synthetic data and preprocessing techniques to improve emotion classification in contexts with limited and imbalanced data.