Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads

Zubiaga et al. (2016c) concluded that users tend to support unverified rumours early on, then shift toward supporting true rumours and debunking false ones over time.

“This study concludes that, while the overall tendency is for users to support unverified rumours in the early stages, there is a shift toward supporting true rumours and debunking false rumours as time goes on.”

Rumours can be classified by various factors, including eventual veracity and degree of credibility.

“Many different factors are available for classifying rumours by type, including its eventual veracity value (true, false, or unresolved) (Zubiaga et al. 2016c) or its degree of credibility (e.g., high or low) (Jaeger et al. 1980).”

Bottom-up sampling strategies emerged more recently to collect a wider range of rumours by sifting through data to identify rumours rather than using known ones.

“Bottom-up sampling strategies have emerged more recently in studies that aimed at collecting a wider range of rumours, i.e., sifting through data to identify rumours, rather than rumours that are already known.”

Veracity annotation has sometimes been operationalised by enlisting journalists with verification expertise.

“This annotation process has been sometimes operationalised by enlisting the help of journalists with expertise in verification (Zubiaga et al. 2016c).”

In one example, journalists analysed rumourous claims during breaking news to determine if a rumour was later confirmed true or debunked false; some rumours were marked unverified because no reliable resolution was found.

“journalists analysed rumourous claims spreading on social media during breaking news to determine, where possible, if a rumour had later been confirmed as true or debunked and proven false; this is, however, not always possible and some rumours were marked as unverified as no reliable resolution could be found.”

Zubiaga et al. (2016c) annotated rumourous tweets for certainty, evidentiality, and support.

“For example, Zubiaga et al. (2016c) annotated rumourous tweets for certainty (certain, somewhat certain, uncertain) and evidentiality (first-hand experience, inclusion of URL, quotation of person/organisation, link to an image, quotation of unverifiable source, employment of reasoning, no evidence), along with support.”

The dataset by Zubiaga et al. (2016) provides over 4,500 tweets categorised by rumour, but unlike Qazvinian's it does not include unrelated tweets with similar characteristics.

“While not specifically intended for rumour tracking, the dataset produced by Zubiaga et al. (2016) provides over 4,500 tweets categorised by rumour.This dataset is different as it does not include tweets with similar characteristics that are actually unrelated.”