Compilation and Validation of a Large Fake News Dataset in Hungarian

Automatically detecting fake news and drawing readers’ attention to possible deceptive intentions can be considered rather useful in the online era. While determining the veracity of writings should involve fact-checking the information contained in the article, using lexical and syntactic clues, text complexity measures, etc. one can predict truthfulness with surprisingly high accuracy. To help the scientific community in studying the fake news phenomenon, we compiled a large dataset of 80 547 legitimate and 67 547 fake news articles from trusted and deceptive Hungarian web portals. The dataset is validated by conducting text classification experiments on the news stories, using traditional bag-of-words and more recent neural network models, the best-performing method achieving a 0.98 F1-score.