An empiric validation of linguistic features in machine learning models for fake news detection

The diffusion of fake news is a growing problem with a high and negative social impact. There are several approaches to address the detection of fake news. This work focuses on a hybrid approach based on functional linguistic features and machine learning. There are several recent works with this approach. However, there are no clear guidelines on which linguistic features are most appropriate nor how to justify their use. Furthermore, many classification results are modest compared to recent advances in natural language processing. Our proposal considers 88 features organized in surface information, part of speech, discursive characteristics, and readability indices. On a 42 677 news database, we show that the classification results outperform previous work, even outperforming state-of-the-art techniques such as BERT, reaching 99.99% accuracy. A proper selection of linguistic features is crucial for interpretability as well as the performance of the models. In this sense, our proposal contributes to the intentional selection of linguistic features, overcoming current technical issues. We identified 32 features that show differences between the type of news. The results are highly competitive in the classification and simple to implement and interpret.