Scholarlink

This study investigates the effectiveness of two transformer-based language models, BERTimbau Base and BERTimbau Large, in classifying students’ affective mood states from forum posts in a Brazilian virtual learning environment. A manually annotated dataset with 5,440 samples across 17 emotional classes was used for training and evaluation. BERTimbau Large achieved 86.89% accuracy, 87.16% precision, 86.89% recall, and 86.86% F1-score, while BERTimbau Base reached 85.66% accuracy, 86.04% precision, 85.66% recall, and 85.69% F1-score. Both models showed difficulty distinguishing semantically similar emotions such as “excited hopeful” and “excited surprised.” Despite this, the results confirm that BERTimbau-based models are effective tools for emotion-aware analysis in educational settings. These findings support their potential for enhancing virtual learning environments through personalized pedagogical interventions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sentiment Analysis in Forums of a Learning Platform: A Comparative Study Between BERTimbau Base and BERTimbau Large Models Regarding Affective Mood States

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Andrio dos Santos Pinto,
  • Magalí Teresinha Longhi,
  • Patricia Alejandra Behar,
  • Jacqueline Mayumi Akazaki

摘要

This study investigates the effectiveness of two transformer-based language models, BERTimbau Base and BERTimbau Large, in classifying students’ affective mood states from forum posts in a Brazilian virtual learning environment. A manually annotated dataset with 5,440 samples across 17 emotional classes was used for training and evaluation. BERTimbau Large achieved 86.89% accuracy, 87.16% precision, 86.89% recall, and 86.86% F1-score, while BERTimbau Base reached 85.66% accuracy, 86.04% precision, 85.66% recall, and 85.69% F1-score. Both models showed difficulty distinguishing semantically similar emotions such as “excited hopeful” and “excited surprised.” Despite this, the results confirm that BERTimbau-based models are effective tools for emotion-aware analysis in educational settings. These findings support their potential for enhancing virtual learning environments through personalized pedagogical interventions.