Evons: A Dataset for Fake and Real News Virality Analysis and Prediction

Proceedings of the 29th International Conference on Computational Linguistics, pages 3589–3596 October 12–17, 2022. 3589 Evons: A Dataset for Fake and Real News Virality Analysis and Prediction Kriste Krstovski,1,2 Angela Soomin Ryu,1 and Bruce Kogut1 1Columbia Business School, Columbia University 2Data Science Institute, Columbia University {kriste.krstovski,asr2193,bruce.kogut}@columbia.edu Abstract We present a novel collection of news arti- cles originating from fake and real news media sources for the analysis and prediction of news virality. Unlike existing fake news datasets which either contain claims or news article headline and body, in this collection each arti- cle is supported with a Facebook engagement count which we consider as an indicator of the article virality. In addition we also pro- vide the article description and thumbnail im- age with which the article was shared on Face- book. These images were automatically anno- tated with object tags and color attributes. Us- ing cloud based vision analysis tools, thumb- nail images were also analyzed for faces and detected faces were annotated with facial at- tributes. We empirically investigate the use of this collection on an example task of article vi- rality prediction. 1 Introduction Fake news articles are widely spread across so- cial media platforms such as Facebook and Twitter. This is mainly due to the fact that social media is gradually becoming the main source of news con- sumption (Shu et al., 2018). Due to the sharing features that these platforms offer, fake news prop- agate rapidly and their effects resonate and persist across many users (Baly et al., 2018). The wide spread of fake news in social media has lead to the development of automatic fake news detection approaches (Ruchansky et al., 2017; Pérez-Rosas et al., 2018; Nguyen et al., 2019; Zellers et al., 2019), to name a few. Developing fake news detec- tion models require annotated collections of fake and real news articles. Most prior work on the creation and annotation of such collections has fo- cused on this task. Significant number of these collections contain claims fact-checked for verac- ity (Vlachos and Riedel, 2014; Wang, 2017). A recent survey of such collections is provided in Guo et al. (2022). On the other hand there exist collections of fake news articles that contain article headline and body text (Horne et al., 2018; Potthast et al., 2018; Zhou et al., 2020). Given that these and other existing fake news collections were developed mainly for fake news detection they can’t be used for analysing and predicting fake news virality which is the set of tasks of our focus. Recently, Shu et al. (2018) cre- ated FakeNewsNet, a collection of ∼24k news arti- cles labeled as fake or real using the fact-checking websites PolitiFact (PolitiFact, 2017) and Gossip Cop (Gossip Cop, 2021). Articles in this collection are annotated with social engagement information obtained through the Twitter search API. However this collection doesn’t include thumbnail images and article descriptions which, along with the head- lines, are the only sources of information readers are exposed to on social media …

Content truncated to 500 words