Detecção de notícias falsas usando técnicas de aprendizado de máquina e aprendizado profundo

Abstract

This work addresses the growing issue of the dissemination of false information, driven by the rise of social media as the main means of communication. The goal of the study is to develop models capable of detecting fake news using combinations of Natural Language Processing, Machine Learning, and Deep Learning techniques. In this study, textual preprocessing strategies were developed, such as the application of stemming and the removal of stopwords, as well as other procedures to ensure the text remained noise-free. Data representations were used to convert the text into a numerical representation, such as Bag-of-Words (TF-IDF), and different word embedding models, including FastText, Word2Vec, and GloVe. For the experiments, a set of different learning algorithms was employed, including traditional machine learning algorithms such as Logistic Regression, Random Forest, Support Vector Machines (SVM), and the Stochastic Gradient Descent Classifier (SGDC), in addition to the use of deep learning architectures, such as LSTM neural networks and the BERT language model. The experiments were conducted on two datasets: Fake.Br and FakeRecogna, both composed of Portuguese news articles labeled as true or false. The experiment with the BERT model achieved the best performance for both datasets, reaching an F1-Score of 96.30% for Fake.Br and 95.21% for FakeRecogna.