Source-linked AI summary
Fighting the COVID-19 Infodemic: Modeling the Perspective of Journalists, Fact-Checkers, Social Media Platforms, Policy Makers, and the Society
Firoj Alam, Shaden Shaar, Fahim Dalvi, Hassan Sajjad, Alex Nikolov, Hamdy Mubarak, Giovanni Da San Martino, Ahmed Abdelali, Nadir Durrani, Kareem Darwish, Abdulaziz Al-Homaid, Wajdi Zaghouani, Tommaso Caselli, Gijs Danoe, Friso Stolk, Britt Bruntink, Preslav Nakov
TL;DR
The paper addresses COVID-19 disinformation across factuality, harm, and policy-relevant questions by releasing a multilingual, stakeholder-informed dataset of 16K manually annotated tweets. It evaluates transformer models using multilingual, contextual, propagandistic-language, bot, and multitask settings, finding sizable gains from these approaches but no universally best model across languages and tasks.
Problem
Social media has become a major channel for disinformation, including rumors, panic, fake cures, xenophobia, and conspiracy theories, creating a need for fine-grained analysis beyond factuality.
Method
The paper develops a 16K-tweet COVID-19 dataset in Arabic, Bulgarian, Dutch, and English, using stakeholder-informed annotations covering check-worthiness, veracity, societal harmfulness, and calls for policy action.
Results
The experiments show sizable performance gains over BERT from multitask learning, cross-language learning, social-context modeling, and propagandistic-language modeling, while no model performs best universally across languages and tasks.
Takeaways & Limitations
The freely available dataset supports multilingual and multifaceted transformer evaluation for research on the COVID-19 infodemic.
Takeaways & Limitations
The dataset could be maliciously used to unfairly moderate non-malicious text, so human moderation is required to prevent this misuse.
Abstract
from arXiv · showhide
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.
1 Introduction
Social media enables broad information sharing but also facilitates COVID-19 disinformation involving falsehoods, harm, panic, and policy-relevant claims. The paper addresses these interconnected issues with a stakeholder-informed, multilingual annotated dataset and reports gains from advanced modeling approaches.
- Social media has become a major information channel, but its openness also makes it a channel for spreading disinformation.
- COVID-19 tweets can spread rumors, panic, fake cures, xenophobia, racism, prejudices, or conspiracy theories beyond simple factual errors.
- Some tweets also blame authorities or discuss actions relevant to policy makers and the general public.
- Rapidly determining factuality, harmfulness, actionability, and verification needs helps organizations prioritize responses to disinformation.
- The study formulates stakeholder-informed annotation around check-worthiness and veracity, societal harm, and calls for government or policy action.
- The dataset contains 16K manually annotated COVID-19 tweets in Arabic, Bulgarian, Dutch, and English, combining perspectives from multiple stakeholders.
- Multitask learning, cross-language learning, social-context modeling, and propagandistic-language modeling produce sizable gains over contextual text representations such as BERT.
- The data and code are made freely available.
2 Related Work
Prior datasets address fact-checking, rumors, propaganda, COVID-19, credibility, prejudice, and related topics, but commonly differ from this work in source, scope, language, or domain. The paper positions its dataset as a multilingual, COVID-19 tweet resource covering multiple disinformation-related perspectives.
- Fact-Checking: Existing fact-checking datasets are often mined from major fact-checking organizations and contain thousands of claims.
- Fact-Checking: Shared-task datasets commonly provide manual annotations for rumors, propaganda, fact-checking, or related claim-analysis tasks.
- Fact-Checking: Unlike this work, many prior datasets use non-tweet sources, target factuality alone, focus on one language, or do not concern COVID-19.
- Check-Worthiness Estimation: Related research also estimates check-worthiness, particularly for claims in political debates.
- COVID-19 Research: COVID-19 Twitter datasets include unlabeled, automatically labeled, distantly supervised, and manually annotated resources.
- COVID-19 Research: Other COVID-19 studies examine credibility, racial prejudice, fear, situational information, mentions, and stance toward misconceptions.
- COVID-19 Research: The closest cited work annotates false and misleading COVID-19 claims across nine topical categories, whereas this paper covers true and false tweets and more perspectives.
- COVID-19 Research: Compared with the authors’ earlier schema work, this study uses a much larger four-language dataset and presents richer experiments.
3 Dataset
The dataset comprises multilingual COVID-19 tweets collected with targeted keywords and annotated across seven questions covering claims, veracity, harm, verification, and policy relevance. Its statistics show skewed label distributions, substantial annotation complexity, and moderate-to-substantial agreement.
- Data Collection: Tweets were collected in four target languages using COVID-19 keywords from January 2020 through March 2021, then deduplicated, filtered, and sampled by engagement.
- Annotation: The annotation schema asks whether tweets contain factual claims and assesses veracity, harm, verification need, and policy-maker interest.
- Annotation: Three annotators per tweet used language-appropriate expertise, with disagreements resolved through consolidation discussions involving external consolidators.
- Annotation: The examples illustrate annotations combining factuality, harm, public interest, urgency, rumors, attacks on officials, and policy relevance.
- Annotation: Fine-grained multiclass labels can be converted to binary labels by mapping Yes* and No* categories and dropping unsure tweets.
- Annotation: Questions Q2–Q4 support either categorical classification or ordinal-regression interpretations, while the tasks remain usable independently.
- Statistics: The collection contains 4,542 English, 4,966 Arabic, 3,697 Bulgarian, and 2,665 Dutch tweets.
- Statistics: Q1 is distributed 64% Yes and 36% No, while 94% of Q3 examples are judged interesting to the general public.
4 Experimental Setup
The experiments evaluate pretrained language models across binary and multiclass tasks, languages, and training scenarios, while adding multitask, cross-language, social-context, propagandistic-language, and bot-related features. Evaluation uses weighted F1 and majority-class baselines, with class imbalance and multiclass difficulty explicitly considered.
- Models and Tasks: The study evaluates binary and multiclass classification across languages using pretrained Transformer representations and additional learning scenarios.
- Data and Evaluation: Data are preprocessed and split stratifiably into 70% training, 10% development, and 20% testing sets.
- Models and Tasks: The experiments compare pretrained contextual models under binary, multiclass, and multilingual setups.
- Models and Tasks: English uses BERT and RoBERTa, Arabic uses AraBERT, Dutch uses BERTje, and multilingual experiments use mBERT and XLM-r.
- Training: Transformer models are fine-tuned for ten epochs, with ten random-seed reruns and development-set model selection.
- Models and Tasks: FastText embeddings pretrained on Common Crawl provide a static-embedding comparison.
- Multitask Learning: Multitask learning uses interrelated questions so information from Q2–Q4 can improve prediction of Q5.
- Twitter/Propagandistic/Botometer Features: The models incorporate Twitter social-context, propagandistic-language, and bot-related features.
5 Evaluation Results
Across languages and classification settings, pretrained models generally beat majority baselines, but no model was universally best. RoBERTa was strongest for English, while XLM-r was especially strong for the remaining languages; multilingual and context-free models also showed notable behavior.
- Most models outperformed the baseline in English, with RoBERTa best on five of seven tasks and FastText best on the other two.
- Except for highly skewed Q3, all Arabic models beat the baseline; FastText and XLM-r each won three of seven tasks.
- Most Bulgarian models outperformed their baselines, with XLM-r best on four tasks and FastText on the remaining three.
- All Dutch models except FastText on Q6 beat the baseline, while XLM-r won five of seven tasks and FastText won two.
- Multiclass classification was harder than binary classification because tasks contained five to ten classes, producing much lower results, including lower baselines.
- No model was universally best across languages, tasks, and class sizes; RoBERTa was strongest for English and XLM-r for the remaining languages.
- Language-specific AraBERT and BERTje were weaker than multilingual XLM-r, partly because they were base-sized whereas XLM-r was large.
- FastText performed strongly, plausibly because character n-grams suit noisy tweets.
6 Advanced Experiments
Advanced experiments tested multilingual training, social and linguistic features, bot-related signals, and multitask learning. Results were mixed for multilingual training but showed improvements from feature combinations and multitask learning on several tasks.
- 6.1 Multilingual Training: Multilingual mBERT fine-tuning produced small, mixed differences: it helped in about half the cases and hurt in the other half.
- 6.1 Multilingual Training: XLM-r performed worse than the multilingual mBERT setup in these experiments.
- 6.2 Twitter/Propagandistic/Botometer: Combining BERT posteriors with Twitter, propaganda, and botness features through XGBoost often improved results, with botness most useful.
- 6.2 Twitter/Propagandistic/Botometer: The feature combinations ranked in usefulness as botness, propaganda, and then Twitter-object features.
- 6.3 Multitask Learning: Multitask BERT and RoBERTa training on English multiclass Q2–Q5 substantially improved Q2, Q4, and Q5 over single-task training.
- 6.3 Multitask Learning: Multitask performance degraded on Q3, probably because its label distribution was skewed.
7 Conclusion and Future Work
The paper releases a large multilingual COVID-19 tweet dataset combining perspectives from multiple stakeholders and evaluates models across languages and advanced learning settings. Future work targets multimodality, explainability, ordinal multitask learning, and practical deployment.
- Its annotation schema combines the perspectives and interests of journalists, fact-checkers, social media platforms, policymakers, and society.
- The authors report transformer evaluations across languages and advanced experiments involving multilingual training, context, propaganda, bot likelihood, and multitask learning.
- Future work will explore multimodality, explainability, ordinal multitask regression, and practical uses of the data and system.
Ethics Statement
The paper limits dataset distribution and intended use while recognizing annotation subjectivity, bias, and risks of misuse. It recommends human supervision for consequential applications.
- The dataset was collected from public tweets through the Twitter API and distributed only as dehydrated Twitter IDs under Twitter’s terms of use.
- Subjective annotations can introduce unavoidable dataset biases, although clear instructions and a defined schema are intended to reduce them.
- Automatic analysis systems may reduce moderators’ workload, but human supervision is required for intricate cases and to prevent harm.
- The models and dataset may be misused by malicious actors, so potential users are asked to account for misuse risks.
- The dataset is distributed for research purposes only without a commercial-use license, and the authors state that identified biases are unintentional.
A Experimental Setup
The experiments evaluate transformer and FastText models alongside social-context, botness, propaganda, and prediction features. Numerical features are log-transformed, while Boolean and categorical features use direct and one-hot encodings.
- The transformer experiments use a learning rate of 2e-5, 10 epochs, and a maximum sequence length of 128.
- Models and Number of Parameters: The study evaluates BERT-family models, multilingual transformers, and FastText embeddings across the four languages.Models include BERT, RoBERTa, AraBERT, BERTje, mBERT, XLM-R, and FastText.
- Additional experiments combine Twitter-object, Botometer, Tanbih propaganda, and BERT prediction features using XGBoost.
- Experiments ran on an NVIDIA Tesla V100-SXM2-32 GB GPU system, with a BERT-base experiment taking about 40 minutes per question on average.
- Boolean features are encoded directly, categorical features use one-hot representation, and numerical values are transformed as x′ = ln(x + 1).
C Detailed Result by Language
The paper defines a multi-question annotation framework for factuality, public interest, harm, and fact-checking, using contextual guidance and explicit label criteria. Evaluation reporting includes accuracy and macro-F1 in addition to weighted F1.
- Evaluation: The reported classification tables include accuracy, macro-F1, and weighted-F1 measures for English, Arabic, Bulgarian, and Dutch.
- Annotation Procedure: Three annotators per tweet could consult linked media, webpages, images, videos, and thread context, while Q2–Q5 were skipped when Q1 was negative.
- D.1 Verifiable Factual Claim: Annotation begins by determining whether a tweet contains a verifiable factual claim, using evidence such as statistics, examples, predictions, laws, or linked media.
- D.3 Interest to the General Public: Public-interest labels use a five-point Likert scale, with healthcare, political news, and current events identified as generally higher-interest topics.
- D.6 Harmful to Society: Harm annotations assess whether content intends to negatively affect society or specific targets, including through weaponized information or rumors.
E Annotation Agreement
Annotation agreement is reported across all four languages and is described as moderate to substantial for binary and multiclass settings. The paper also provides language-specific label-distribution statistics.
- Agreement Results: Overall inter-annotator agreement was moderate to substantial across all questions in both binary and multi-label settings.
- Agreement Measure: Agreement is quantified with Fleiss Kappa between each annotator and the consolidated label using original multiclass and binary labels.
- Language Comparisons: Tables report inter-annotator agreement separately for English, Arabic, Bulgarian, and Dutch.
- Label Distributions: Figures 3–6 provide detailed manual-annotation label distributions for English, Arabic, Bulgarian, and Dutch.
G.1 English Tweets
Across the English, Arabic, Bulgarian, and Dutch datasets, label relationships show consistent links among factuality, public interest, harmfulness, and fact-checking relevance. Additional analyses connect account verification and multimedia presence with claim, impact, factuality, and harm labels, while English and Arabic tweets show geographically diverse origins.
- Question-pair relationships: In Arabic tweets, “no false info” is associated with public interest at 76%, while public interest is higher for non-harmful tweets at 79% than harmful tweets at 21%.False-information labels are associated with harmful content, whereas claims without false information are often judged unnecessary or too trivial to check.
- Question-pair relationships: The Bulgarian and Dutch heatmaps produce observations very similar to those reported for English and Arabic tweets.The paper presents corresponding contingency and correlation heatmaps for both languages.
- Geographic distribution: Most English tweets originate from the USA, UK, Canada, and India, while most Arabic tweets originate from Gulf countries, with both languages spanning multiple countries.The authors interpret this distribution as providing diversity in interests, topics, and style; Bulgarian and Dutch are not analyzed geographically because speakers are concentrated in Bulgaria and the Netherlands.
- Account verification: Verified accounts more often post factual claims and are less likely than unverified accounts to post false or harmful information.Their tweets are also more likely to interest the general public, and the authors add corresponding account-verification features to their models.
- Media types: Tweets containing multimedia are more likely to include factual claims and have higher public impact, but less likely to contain false information or be harmful.These observations motivate modeling multimedia as a feature.