Early Fake News Detection Using Linguistic Features: A Comparative Study of Machine Learning and Transformer Models

Misinformation and disinformation spread rapidly on social media, so reliable mechanisms are needed to classify news as fake or real. Although deep learning models have become increasingly popular, traditional machine learning methods remain relevant because they are easier to interpret and can be very efficient. This paper proposes a lightweight framework for fake news detection that uses linguistic and sentiment-based features of the textual content. The framework was evaluated in detail, and the results show that the Bidirectional Encoder Representations from Transformers (BERT) model outperformed the other models in accuracy, while traditional models with engineered features performed adequately at a much lower computational cost. The results demonstrate the effectiveness of simple yet powerful methods for detecting fake news from the text of complete news articles.