A large-scale benchmark dataset for rumor detection on Sina Weibo, sourced from the Sina community management center which reports verified rumors, with non-rumors gathered from general threads.
Overview
- Events: 4,664 (2,313 rumors, 2,351 non-rumors)
- Posts: 3,805,656 from 2,746,818 unique users
- Time span: 2012–2015
- Average posts per event: 816
- Maximum posts per event: 59,318
- Average event duration: 2,460.7 hours
Construction
Rumors were obtained from Sina's official community management center (http://service.account.weibo.com), which aggregates and verifies misinformation. Non-rumor events were gathered by crawling general threads not reported as rumors. The Weibo API was used to retrieve original messages and all repost/reply messages for each event.
Characteristics
The Weibo dataset is substantially larger than the Twitter version and exhibits different properties: higher average event duration, larger maximum post counts, and more balanced rumor/non-rumor split. The larger scale and richer conversational structure (repost/reply hierarchy) make it valuable for studying temporal and structural signals in rumor propagation.
Usage
Used to evaluate RNN-based rumor detection models (2016-ma-detecting-rumors-microblogs) and serves as a cross-platform benchmark for comparing detection approaches across English-language (Twitter) and Chinese-language (Weibo) social media.
Download
Available at: http://alt.qcri.org/~wgao/data/rumdect.zip (packaged with the Twitter dataset and code)
Related
- Twitter Rumor Dataset — corresponding English-language dataset from Twitter (992 events)
- Rumor detection on social media