ReCOVery¶
Full name: ReCOVery: A Multimodal Repository for COVID-19 News Credibility Research Authors: Zhou Xinyi, Mulay Apurva, Ferrara Emilio, Zafarani Reza Paper: Zhou et al. (2020), CIKM '20 Access: http://coronavirus-fakenews.com (tweet IDs and article IDs; full text via provided instructions)
Description¶
ReCOVery is a multimodal dataset of COVID-19 news articles labeled for credibility. Labels are assigned at the publisher level using NewsGuard and Media Bias/Fact Check (MBFC) rather than at the per-article level, enabling scalable collection without manual annotation. This design allows continuous extension as publishers release more COVID-19 content.
Statistics¶
| Split | Articles | w/ images | w/ social data | Tweets | Users |
|---|---|---|---|---|---|
| Reliable | 1,364 | 1,354 | 1,219 | 114,402 | 78,659 |
| Unreliable | 665 | 663 | 528 | 26,418 | 17,323 |
| Total | 2,029 | 2,017 | 1,747 | 140,820 | 93,761 |
Class ratio: approximately 2:1 reliable to unreliable.
Schema¶
Each news article has 12 components:
| Field | Description |
|---|---|
| News ID | Unique article identifier |
| URL | Source URL |
| Publisher | Name of the publishing media outlet |
| Publication date | yyyy-mm-dd format |
| Author(s) | Byline (may be blank or fictional) |
| Title | News headline |
| Body text | Full article text |
| Main image | URL of the primary/head image |
| Country | Country of the publisher |
| Political bias | One of: extremely left / left / left-center / center / right-center / right / extremely right (from AllSides + MBFC) |
| NewsGuard score | 0–100 credibility score |
| MBFC factual level | very high / high / most factual / mixed / low / very low |
Social data fields: tweet ID, tweet text, language, creation time, retweet/reply/like counts; posting user ID, follower count, friend count.
Labeling methodology¶
Publishers are classified using strict thresholds: - Reliable: NewsGuard score >90 AND MBFC factual level = "very high" or "high" - Unreliable: NewsGuard score <30 AND MBFC factual level = "low" or "very low"
Threshold of 90/30 (vs. NewsGuard's default 60) is intended to reduce false positives/negatives. 22 reliable publishers (e.g., NPR, Reuters) and 38 unreliable publishers (e.g., Humans Are Free, Natural News) are included. Coverage: US, Russia, UK, Iran, Cyprus, Canada.
Intended use¶
- Credibility classification of COVID-19 news
- Multimodal fake news detection (text + image + social)
- Analysis of misinformation propagation dynamics during a pandemic
- Study of political bias in COVID-19 news coverage
Limitations¶
Publisher-level labels introduce noise: a credible publisher may occasionally publish an error; an unreliable publisher may republish a factual story. The corpus is US-heavy and English-language only. Tweet data is released as IDs only; historical tweet retrieval may be incomplete due to tweet deletions.
Connections¶
- FakeNewsNet is the closest structural analogue (news + Twitter spreading), covering general fake news with per-article fact-check labels rather than publisher-level labels.
- SAFE (Zhou et al., 2020) is evaluated on ReCOVery as a multimodal baseline and achieves the best benchmark F₁.
- Cao et al. (2025) — SLIM uses ReCOVery (train/val/test as provided by the dataset, 966/278/120 reliable and 487/114/64 fake) as a primary benchmark for limited-information detection; reports 95.55% accuracy with the full-text XLNet baseline and ~99% accuracy ratio with 30% keyword extraction.