COVID-19 Misinformation Detection: Machine-Learned Solutions to the Infodemic.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Public Health

Background:

• The proliferation of COVID-19 misinformation overwhelms human fact-checking capabilities.
• Automated and machine learning (ML) methods offer scalable solutions for online misinformation detection.
• There is an urgent need for improved ML-based strategies to combat health infodemics.

Purpose of the Study:

• To enhance automated and machine-learned methods for responding to health-related misinformation.
• To evaluate the performance of ML models trained on different datasets for misinformation classification.

Main Methods:

• Trained ML models using three strategies: COVID-19 data only, general fact-checked data only, and combined data.
• Created two COVID-19 misinformation datasets (~7,000 and ~31,000 entries) from verified and programmatic sources.
• Utilized crowdsourced human votes (31,441) to label one dataset for model training and validation.

Main Results:

• Models achieved high external validation accuracies: 96.55% and 94.56% on two separate datasets.
• The best performance was achieved using a model fine-tuned on COVID-19-specific content.
• Blended models combining ML predictions with human votes reached up to 99.1% accuracy, significantly outperforming human votes alone (73%).

Conclusions:

• ML models demonstrate superior performance in classifying COVID-19 content veracity.
• Fine-tuning pretrained language models on topic-specific data yields optimal results.
• Crowdsourced data and blended human-ML approaches enhance accuracy, offering a robust strategy against health disinformation.