Automatic health misinformation detection on social media: an NLP-driven classifier for combatting pseudoscientific and unverified claims

By combining cutting-edge natural language processing (NLP) methods with machine learning models, this study tackles the crucial problem of identifying COVID-19-related health disinformation on social media platforms. Unlike prior studies, this work emphasizes the effective combination of transformer-based models (DistilBERT, RoBERTa, CT-BERT) with traditional classifiers (LightGBM, Logistic Regression, and XGBoost) in a cohesive pipeline. A key contribution is the development of a robust preprocessing and data-balancing strategy using the Synthetic Minority Over-sampling Technique (SMOTE), which mitigates the impact of class imbalance in sentiment labels, enhancing model reliability. The study also explores the role of domain-specific fine-tuning of transformers, achieving superior accuracy while reducing computational costs compared to larger models. Results demonstrate that CT-BERT achieves state-of-the-art performance with a 97% accuracy rate, validating the efficacy of contextual language understanding in misinformation detection. This novel methodology provides actionable insights for real-world applications such as public health monitoring, misinformation tracking, and evidence-based decision-making.