Scholarlink

This study explores the integration of Large Language Models with traditional machine learning techniques to classify COVID-19 related tweets, leveraging the lightweight Mistral model developed by Mistral AI. Enhanced by word2vec-based data augmentation, the approach generates dynamic, class-defining attributes to produce high-quality, contextually relevant text. This dual-model framework combines Mistral’s preprocessing strengths with BERT and a Random Forest classifier to effectively address themes such as vaccines, masks, quarantine, and social distancing. This exploratory case study demonstrates that LLM labeling and strategic data augmentation can significantly improve accuracy on small datasets and provide a scalable solution for social media content analysis. The BERT model achieved 94% accuracy with simple augmentation and 91% accuracy with advanced augmentation, while the Random Forest model showed lower performance, classifying fewer examples correctly. The study highlights the effectiveness of LLM-generated labels and advanced data augmentation, particularly with the BERT model, in enhancing classification accuracy, semantic relevance, and reducing uncertainty. Future research will focus on expanding the range of data classification and improving the quality of dynamically generated class attributes to better capture semantic complexity and further enhance model performance.

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

Leveraging Data Augmentation and Large Language Models for Enhanced COVID-19 Tweet Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Eric Landaverde,
  • Adam Spencer,
  • Daehan Kwak

摘要

This study explores the integration of Large Language Models with traditional machine learning techniques to classify COVID-19 related tweets, leveraging the lightweight Mistral model developed by Mistral AI. Enhanced by word2vec-based data augmentation, the approach generates dynamic, class-defining attributes to produce high-quality, contextually relevant text. This dual-model framework combines Mistral’s preprocessing strengths with BERT and a Random Forest classifier to effectively address themes such as vaccines, masks, quarantine, and social distancing. This exploratory case study demonstrates that LLM labeling and strategic data augmentation can significantly improve accuracy on small datasets and provide a scalable solution for social media content analysis. The BERT model achieved 94% accuracy with simple augmentation and 91% accuracy with advanced augmentation, while the Random Forest model showed lower performance, classifying fewer examples correctly. The study highlights the effectiveness of LLM-generated labels and advanced data augmentation, particularly with the BERT model, in enhancing classification accuracy, semantic relevance, and reducing uncertainty. Future research will focus on expanding the range of data classification and improving the quality of dynamically generated class attributes to better capture semantic complexity and further enhance model performance.