Scholarlink

<p>The COVID-19 pandemic triggered an unprecedented surge in Twitter activity, providing a rich source of public opinions and emotions. This study proposes a fusion model combining a bidirectional GRU (BiGRU) and a DistilBERT transformer, with their learned features concatenated and fed into an XGBoost meta-classifier for final sentiment prediction. We evaluate our approach on over one million COVID-19-related English-language tweets collected from eight countries between January and April 2020. The fusion model (BiGRU + DistilBERT + XGBoost) achieves a classification accuracy of ~ 85.8%, outperforming individual models (for example, DistilBERT alone at 85.5% accuracy) in sentiment detection. Key results indicate that public sentiment evolved with pandemic phases: negative sentiments peaked during surges in cases and deaths, while positive sentiments rose during recovery periods. We also observe notable regional differences – for instance, sentiment trends vary across countries and between English- vs. non-English-speaking regions. These findings demonstrate the effectiveness of our context-infused approach, offering valuable insights for policymakers on social media sentiment during health crises.</p>

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

Context-based sentiment analysis using a BiGRU DistilBERT fusion model for COVID-19 tweets

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月11日

查看全文
  • Utkarsh Sharma,
  • Prateek Pandey,
  • Shishir Kumar

摘要

The COVID-19 pandemic triggered an unprecedented surge in Twitter activity, providing a rich source of public opinions and emotions. This study proposes a fusion model combining a bidirectional GRU (BiGRU) and a DistilBERT transformer, with their learned features concatenated and fed into an XGBoost meta-classifier for final sentiment prediction. We evaluate our approach on over one million COVID-19-related English-language tweets collected from eight countries between January and April 2020. The fusion model (BiGRU + DistilBERT + XGBoost) achieves a classification accuracy of ~ 85.8%, outperforming individual models (for example, DistilBERT alone at 85.5% accuracy) in sentiment detection. Key results indicate that public sentiment evolved with pandemic phases: negative sentiments peaked during surges in cases and deaths, while positive sentiments rose during recovery periods. We also observe notable regional differences – for instance, sentiment trends vary across countries and between English- vs. non-English-speaking regions. These findings demonstrate the effectiveness of our context-infused approach, offering valuable insights for policymakers on social media sentiment during health crises.