Analysis of Text Feature Extractors using Deep Learning on Fake News

Abstract

Social media and easy internet access have allowed the instant sharing of news, ideas, and information on a global scale. However, rapid spread and instant access to information/news can also enable rumors or fake news to spread very easily and rapidly. In order to monitor and minimize the spread of fake news in the digital community, fake news detection using Natural Language Processing (NLP) has attracted significant attention. In NLP, different text feature extractors and word embeddings are used to process the text data. The aim of this paper is to analyze the performance of a fake news detection model based on neural networks using 3 feature extractors: TD-IDF vectorizer, Glove embeddings, and BERT embeddings. For the evaluation, multiple metrics, namely accuracy, precision, F1, recall, AUC ROC, and AUC PR were computed for each feature extractor. All the transformation techniques were fed to the deep learning model. It was found that BERT embeddings for text transformation delivered the best performance. TD-IDF has been performed far better than Glove and competed the BERT as well at some stages.