Rating news claims: Feature selection and evaluation

News claims that travel the Internet and online social networks (OSNs) originate from different, sometimes unknown sources, which raises issues related to the credibility of those claims and the drivers behind them. Fact-checking websites such as Snopes, FactCheck, and Emergent use human evaluators to investigate and label news claims, but the process is labor- and time-intensive. Driven by the need to use data analytics and algorithms in assessing the credibility of news claims, we focus on what can be generalized about evaluating human-labeled claims. We developed tools to extract claims from Snopes and Emergent and used public datasets collected by and published on those websites. Claims extracted from those datasets were supervised or labeled with different claim ratings. We focus on claims with definite ratings-false, mostly false, true, and mostly true, with the goal of identifying distinctive features that can be used to distinguish true from false claims. Ultimately, those features can be used to predict future unsupervised or unlabeled claims. We evaluate different methods to extract features as well as different sets of features and their ability to predict the correct claim label. By far, we noticed that OSN websites report high rates of false claims in comparison with most of the other website categories. The rate of reported false claims is higher than the rate of true claims in fact-checking websites in most categories. At the content-analysis level, false claims tend to have more negative tones in sentiments and hence can provide supporting features to predict claim classification.