Aplicação de PLN para classificação de comentários de jogos em língua portuguesa.

Abstract

This work investigates the application of Natural Language Processing (NLP) techniques to the multi-label classification of Portuguese-language digital game reviews, in the context of automatic user opinion analysis. The proposed approach consists of a hybrid pipeline that combines topic modeling (Latent Dirichlet Allocation), automatic labeling using Large Language Models (LLMs), and consensus strategies among multiple models to improve label consistency. Based on this dataset, supervised learning models are evaluated under different configurations of text preprocessing and vector representations. Initially, real user comments were collected and processed using text normalization and cleaning techniques, and subsequently enriched with automatically generated labels, reducing the dependence on manual annotation. Then, classification algorithms such as Support Vector Machines (SVM), Random Forest, and Logistic Regression were applied in multi-label settings. The results indicate significant performance variations across the evaluated configurations, highlighting the impact of preprocessing strategies and the quality of automatically generated labels. Overall, the integration of supervised learning with modern data construction methods shows potential for performance improvement, although challenges remain regarding label consistency and model generalization. As future work, we propose expanding the dataset, adopting more robust consensus strategies among LLMs, and incorporating additional supervised learning models.