Automated Fake News Detection Using Computational Forensic Linguistics

Fake news is news-like content that has been produced without following journalism principles. Fake news try to mimic the look and feel of real news to intentionally disinform the reader. This phenomenon can have a strong influence on society, thus being potentially a severe problem. To address this phenomenon, systems to detect fake news have been developed, but most of them build upon fact-checking approaches, which are unfit to detect misinformation when a news piece, rather than completely false, is distorted, exaggerated, or even decontextualized. We aim to detect Portuguese fake news by following a forensic linguistics approach. Contrary to previous approaches, we build upon methods of linguistic and stylistic analysis that have been tried and tested in forensic linguists. After collecting corpora from multiple fake news outlets and from a genuine news source, we formulate the task as a text classification problem and demonstrate the effectiveness of the proposed features when training different classifiers for telling fake from genuine news. Furthermore, we perform an ablation study with subsets of features and find that the proposed feature sets are complementary. The highest results reported are very promising, achieving 97% of accuracy and a macro F1-score of 91%.