Fighting Health-Related Misinformation In Social Media With Large Language Models.

Area of Science:

• Artificial Intelligence
• Public Health
• Computational Linguistics

Background:

• Social media platforms face significant challenges with the spread of health disinformation.
• Accurate identification and rebuttal of medical misinformation are crucial for public health.
• Existing methods for combating online health misinformation require enhancement.

Purpose of the Study:

• To fine-tune Large Language Models (LLMs) for identifying and countering health-related disinformation on social media.
• To evaluate the performance of T5, BERT, and LlaMa-2 models in classifying health content and detecting disinformation.
• To develop a system utilizing Retrieval Augmented Generation (RAG) for rebutting identified misinformation.

Main Methods:

• Fine-tuning T5, BERT, and LlaMa-2 models in two stages: health-related text classification and disinformation verification.
• Employing Retrieval Augmented Generation (RAG) to query trusted medical databases for accurate information.
• Experimentally assessing model performance using precision, recall, and F1 scores.

Main Results:

• Models achieved 94% precision, 95% recall, and 90% F1 for classifying health-related content.
• Disinformation texts were classified with 99% precision, 95% recall, and 97% F1.
• The developed system demonstrates high accuracy in identifying and classifying health disinformation.

Conclusions:

• Fine-tuned LLMs are highly effective tools for combating health disinformation on social media.
• The RAG approach enables efficient rebuttal of misinformation by leveraging authoritative medical sources.
• The presented system offers a valuable solution for health experts to manage online health-related disinformation.