Abstract
To evaluate credibility of Arabic content in Twitter, a corpus of Arabic microblogging messages was required with its labelled credibility ratings in order to build the credibility model. Since no Arabic dataset existed, we confronted this problem by building a novel human annotated Arabic Twitter corpus that could be used for further research. This paper identifies the collection process and the characteristics of the newly created dataset. It presents basic analysis of submitted credibility rating values and the collected labelers' data. A number of statistical graphs are exhibited to examine labelers' traits and its impact on credibility perceptions. Results showed that both: labelers' data and method of labeling presentation have a slight impact on the perception of credibility. The results presented in this paper covers the first stage from a large project aims at predicting credibility of Arabic Twitter messages in the presence of disagreed judging credibility scores.