Multi-label multi-class COVID-19 Arabic Twitter dataset with fine-grained misinformation and situational information...

Area of Science:

• Computational Social Science
• Natural Language Processing
• Public Health Informatics

Background:

• The COVID-19 pandemic saw a surge in social media misinformation, impacting public health and vaccination efforts.
• Existing datasets often lack fine-grained classification for misinformation types and situational context.
• Accurate identification of misinformation sub-types is crucial for targeted interventions and public safety.

Purpose of the Study:

• To develop comprehensive annotation guidelines for fine-grained misinformation and situational information.
• To release the first Arabic COVID-19 misinformation dataset with multi-class and multi-label annotations.
• To establish baseline performance for misinformation and situational information classification models.

Main Methods:

• Developed 19 fine-grained misinformation classes and 6 situational information classes.
• Created an Arabic dataset of approximately 6.7K tweets with multi-class/multi-label annotations.
• Experimented with machine learning and transformer-based classifiers (including AraBERT-COV19) for classification tasks.

Main Results:

• The dataset is validated as suitable for building misinformation and situational information classification models.
• AraBERT-COV19 achieved high performance (81.6% F-score) for multi-class misinformation classification.
• Label Powerset with linear SVC demonstrated strong performance (76.69% F-score) for multi-label misinformation classification.

Conclusions:

• The developed dataset and classification models are vital for combating COVID-19 misinformation in Arabic.
• Fine-grained classification enables more effective strategies for managing online health information.
• This work provides a foundation for future research in Arabic misinformation detection and emergency response.