A benchmark dataset for rumor detection on Twitter, constructed by confirming rumors and non-rumors from Snopes.com and augmenting with additional non-rumors from prior datasets to balance the classes.
Overview
- Events: 992 (498 rumors, 494 non-rumors)
- Posts: 1,101,985 from 491,229 unique users
- Time span: March–December 2015
- Average posts per event: 1,111
- Source: Snopes.com with refined keyword queries; non-rumor events from 2011-castillo-twitter-credibility and 2013-kwon-features-rumor-propagation
Construction
For each event, authors extracted keywords from the Snopes URL (e.g., http://www.snopes.com/pentagon-spends-powerball-tickets → keywords: pentagon, spends, powerball, tickets), refined them manually, and iteratively queried Twitter's "Live" search API to retrieve matching posts. The resulting collection was balanced to 50/50 rumors/non-rumors using prior public datasets.
Usage
Used to evaluate RNN-based rumor detection models (2016-ma-detecting-rumors-microblogs) and serves as a benchmark for comparing deep learning and hand-crafted feature approaches.
Download
Available at: http://alt.qcri.org/~wgao/data/rumdect.zip (packaged with the Weibo dataset and code)
Related
- Weibo Rumor Dataset — corresponding Chinese-language dataset from Sina Weibo (4,664 events)
- Rumor detection on social media