Análise de sentimentos em comentários do youtube: uma investigação sobre música popular contemporânea

Abstract

This study investigates the sentiments expressed by users in comments posted on contemporary pop music videos on YouTube. With the growth of digital platforms and the increasing interaction between users and audiovisual content, it has become possible to systematically observe how listeners react to the music they consume. To this end, this research employs the Gemma 3 4B language model using a few-shot approach to classify comment polarity into three categories: positive, negative, and neutral. Approximately 21,700 comments were collected, of which 4,705 were manually annotated and used as a reference set for evaluating the automatic classification. The model’s performance was assessed using accuracy, precision, recall, and F1-score. In the global evaluation, the model achieved an accuracy of 0.849 and a macro F1-score of 0.826. In the per-video analysis, macro F1-scores ranged from 0.538 to 0.943, demonstrating performance differences across the analyzed comment sets. The qualitative inspection identified examples of mixed evaluations and context-dependent expressions that may guide further investigation of classification errors, without establishing a causal relationship between these characteristics and model performance. The results indicate that Gemma 3 4B has potential for automatic sentiment classification in music video comments, although its performance varies across different comment sets