TweetCred: A Real-time Web-based System for Assessing Credibility of Content on Twitter

During large scale events, a large volume of content is posted on Twitter, but not all of this content is trustworthy. The presence of spam, advertisements, rumors and fake images reduces the value of information collected from Twitter, especially during sudden-onset crisis events where information from other sources is scarce. In this research work, we describe various facets of assessing the credibility of user-generated content on Twitter during large scale events, and develop a novel real-time system to assess the credibility of tweets. Firstly, we develop a semi-supervised ranking model using SVM-rank for assessing credibility, based on training data obtained from six high-impact crisis events of 2013. An extensive set of forty-five features is used to determine the credibility score for each of the tweets. Secondly, we develop and deploy a system--TweetCred--in the form of a browser extension, a web application and an API at the link: http://twitdigest.iiitd.edu.in/TweetCred/. To the best of our knowledge, this is the first research work to develop a practical system for credibility on Twitter and evaluate it with real users. TweetCred was installed and used by 717 Twitter users within a span of three weeks. During this period, a credibility score was computed for more than 1.1 million unique tweets. Thirdly, we evaluated the real-time performance of TweetCred, observing that 84% of the credibility scores were displayed within 6 seconds. We report on the positive feedback that we received from the system's users and the insights we gained into improving the system for future iterations.