Scholarlink

In this study, we developed an Emotion Detection (ED) model for Brazilian Portuguese social media comments and used it to analyze emotion distribution in replies to COVID-19-related news on Twitter. We translated the GoEmotions dataset into Portuguese using a Large Language Model, preserving the original social media writing style, and fine-tuned the BERTimbau model for detecting Ekman’s six basic emotions plus neutral emotion. We collected tweets related to the COVID-19 pandemic from major Brazilian news portals and analyzed user replies for expressed emotions. Chi-square tests indicated that emotion distribution depended on both the news topic and the media outlet, while the Augmented Dickey-Fuller (ADF) test showed that some emotions’ distributions were influenced by the publication date. Among replies with identifiable emotions, anger (21.1%), joy (17.2%), and surprise (11.6%) were most prevalent. The Brazilian pandemic inquiry (CPI) prompted the highest anger prevalence (41.9%), while news from UOL Notícias had the highest anger (40.0%) and g1 led in sadness (13.6%).

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

Portuguese Emotion Detection Model Using BERTimbau Applied to COVID-19 News and Replies

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Francisco Bráulio Oliveira,
  • Jaime Simão Sichman

摘要

In this study, we developed an Emotion Detection (ED) model for Brazilian Portuguese social media comments and used it to analyze emotion distribution in replies to COVID-19-related news on Twitter. We translated the GoEmotions dataset into Portuguese using a Large Language Model, preserving the original social media writing style, and fine-tuned the BERTimbau model for detecting Ekman’s six basic emotions plus neutral emotion. We collected tweets related to the COVID-19 pandemic from major Brazilian news portals and analyzed user replies for expressed emotions. Chi-square tests indicated that emotion distribution depended on both the news topic and the media outlet, while the Augmented Dickey-Fuller (ADF) test showed that some emotions’ distributions were influenced by the publication date. Among replies with identifiable emotions, anger (21.1%), joy (17.2%), and surprise (11.6%) were most prevalent. The Brazilian pandemic inquiry (CPI) prompted the highest anger prevalence (41.9%), while news from UOL Notícias had the highest anger (40.0%) and g1 led in sadness (13.6%).