Detecção de Fake News em Português a partir de poucos dados rotulados

Abstract

There have been studies on detecting false news for several years, but most studies are based on news produced in the English language. This work presents a study to detect false news in news produced in Brazilian Portuguese language. The study shows an experimental analysis of the behavior of supervised and semi-supervised classifiers capable of predicting whether or not news is categorized as false news. Concepts that a dataset has most of the data without a class label and that this data may or may not be false news, these definitions identify a new category of classifiers trained from positive examples and unlabeled examples, or, PU (Positive and Unlabeled) learning. The results of the study show that PU learning outperforms other supervised and semi-supervised learning classifiers, as it can effectively identify more fake news in a scenario where there are few positive examples.