Alphabet Flatting as a variant of n-gram feature extraction method in ensemble classification of fake news

Abstract

The detection of disinformation becomes a significant challenge in the modern world. Most of our communication media and most of the sources of information about reality are located on the distributed network services, where the published content is usually not a subject to any initial verification. One of the few tools that seem to be able to process such large volumes of data efficiently are pattern recognition methods employing extraction of features obtained through the Natural Language Processing models and procedures. The following paper is proposing an Alphabet Flatting – a modification of the preprocessing method for the feature extraction from large language corpora – allowing the construction of diverse classifier ensembles integrated by the support accumulation, the generalization power of which may compete with quality of the state-of-the-art models in environments with strict time constraints. The proposed method has been thoroughly evaluated with the set of computer experiments, the results of which allow us to conclude its potential usefulness in the solutions of the automatic systems for preventing the spread of fake news.