Fighting the COVID-19 Infodemic: Modeling the Perspective of Journalists, Fact-Checkers, Social Media Platforms, Policy Makers, and the Society

Firoj Alam, Shaden Shaar, Alex Nikolov, Hamdy Mubarak, Giovanni Da San Martino, Ahmed Abdelali, Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Kareem Darwish, Preslav Nakov

TL;DR

This paper tackles the COVID-19 infodemic by releasing a large, multi-lingual dataset of 16,000 annotated tweets. They developed a fine-grained annotation schema that captures multiple perspectives (journalists, fact-checkers, policymakers, etc.) to model disinformation, demonstrating the effectiveness of transformer models on this complex task.

What Problem Does This Paper Address?

The COVID-19 pandemic saw an unprecedented surge in misinformation, dubbed the "infodemic." This wasn't just about factual inaccuracies; it encompassed rumors, fake cures, conspiracy theories, and even xenophobia, posing significant public health and societal risks. Existing efforts to combat this often focused on isolated aspects, like just fact-checking claims. However, understanding the full picture requires considering various facets: is a claim verifiable? Is it false? Is it harmful? Is it interesting to the public or policymakers? This paper aims to address this by creating a comprehensive dataset and framework to analyze COVID-19 disinformation from multiple stakeholder viewpoints.

Methods

The core of this work lies in its novel annotation schema and the subsequent modeling. They developed a multi-faceted approach to annotate tweets, asking seven key questions:

  1. Verifiable Factual Claim: Does the tweet contain a claim that can be fact-checked?
  2. False Information: To what extent does the tweet appear false? (5-point Likert scale)
  3. General Public Interest: Is the claim of interest to the general public? (5-point Likert scale)
  4. Harmfulness: To what extent is the tweet harmful to society, individuals, or products? (5-point Likert scale)
  5. Need for Verification: Should a professional fact-checker verify the claim? (Categorical)
  6. Harm to Society: Is the tweet harmful to society and why? (Categorical with specific reasons like rumor, panic, xenophobia, etc.)
  7. Attention from Government Entities: Should the tweet get attention from policymakers? (Categorical with specific reasons like blaming authorities, containing advice, calling for action, etc.)

They then employed state-of-the-art transformer models (BERT, RoBERTa, XLM-R) for classification, exploring monolingual, multilingual, single-task, and multitask learning settings. They also incorporated Twitter-specific features (user context, botness, propaganda scores) to enhance model performance.

Datasets

The primary contribution is the introduction of a new dataset: 16,000 manually annotated tweets related to COVID-19. This dataset is unique in that it:

  • Focuses specifically on the COVID-19 infodemic.
  • Combines perspectives from journalists, fact-checkers, social media platforms, policymakers, and society.
  • Covers four languages: Arabic, Bulgarian, Dutch, and English.
  • Provides fine-grained annotations for multiple aspects of disinformation.

The paper also references existing datasets in related work, such as Liar, ClaimsKG, and MultiFC, but emphasizes how their dataset is more comprehensive and COVID-19-specific.

Key Insights and Ideas

  • Multifaceted Disinformation: The "infodemic" is not just about factual errors. Tweets can be harmful, spread panic, or be of public interest even if factually correct. This necessitates a nuanced approach beyond simple true/false classification.
  • Stakeholder Perspectives Matter: Incorporating the viewpoints of different actors (journalists, policymakers, etc.) leads to a richer understanding of disinformation and its impact.
  • Transformer Models Excel: Pre-trained transformer models, especially multilingual ones like XLM-R, show strong performance across various languages and tasks, demonstrating their utility for tackling complex NLP challenges like disinformation detection.
  • Multitask Learning Benefits: Combining related tasks (e.g., predicting harm and public interest) can improve performance on specific questions, suggesting that these aspects are interconnected.
  • Social and Contextual Features are Crucial: Incorporating features like botness, propaganda scores, and user context significantly boosts model performance, highlighting the importance of understanding the social dynamics of information spread.

Key Results

  • The proposed models achieved strong performance across all four languages and multiple tasks, often outperforming majority class baselines.
  • For English, RoBERTa was the most successful model in five out of seven tasks.
  • For Arabic, FastText and XLM-r each won three tasks.
  • XLM-r was particularly strong for Bulgarian and Dutch, winning four and five tasks respectively.
  • Multitask learning showed improvements for certain questions (Q2, Q4, Q5) compared to single-task setups.
  • The inclusion of Twitter-specific features (botness, propaganda) led to notable improvements.
  • Verified accounts tend to post more factual claims, less false information, and their tweets are more likely to be of public interest and less harmful.

Limitations and Future Work

  • Subjectivity in Annotations: The authors acknowledge that some annotations are subjective, and biases are inevitable. They emphasize the clear annotation schema and instructions to mitigate this.
  • Language-Specific Models vs. Multilingual: While language-specific models (AraBERT, BERTje) were tested, multilingual models like XLM-r often performed better, possibly due to their larger size and broader training data.
  • Future Work: The authors plan to explore multimodal analysis (images, videos) and explainability. They also aim to model the task as a multitask ordinal regression and to deploy the system for practical use, potentially in future shared tasks.

Summary updated

Abstract

With the emergence of the COVID-19 pandemic, the political and the medical aspects of disinformation merged as the problem got elevated to a whole new level to become the first global infodemic. Fighting this infodemic has been declared one of the most important focus areas of the World Health Organization, with dangers ranging from promoting fake cures, rumors, and conspiracy theories to spreading xenophobia and panic. Addressing the issue requires solving a number of challenging problems such as identifying messages containing claims, determining their check-worthiness and factuality, and their potential to do harm as well as the nature of that harm, to mention just a few. To address this gap, we release a large dataset of 16K manually annotated tweets for fine-grained disinformation analysis that (i) focuses on COVID-19, (ii) combines the perspectives and the interests of journalists, fact-checkers, social media platforms, policy makers, and society, and (iii) covers Arabic, Bulgarian, Dutch, and English. Finally, we show strong evaluation results using pretrained Transformers, thus confirming the practical utility of the dataset in monolingual vs. multilingual, and single task vs. multitask settings.
BibTeX
@inproceedings{alam2021fighting,
  author = {Firoj Alam and Shaden Shaar and Alex Nikolov and Hamdy Mubarak and Giovanni Da San Martino and Ahmed Abdelali and Fahim Dalvi and Nadir Durrani and Hassan Sajjad and Kareem Darwish and Preslav Nakov},
  title = {{Fighting the COVID-19 Infodemic: Modeling the Perspective of Journalists, Fact-Checkers, Social Media Platforms, Policy Makers, and the Society}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.findings-emnlp.56},
}
Figure