A Multi-Model Approach to Detecting Clickbait and Fake News Using BERT Embeddings and Topic Modeling

With online information dissemination widespread in this age, separating fake news from clickbait is an essential challenge. This study offers a new hybrid approach that combines Bidirectional Encoder Representations from Transformers (BERT) embeddings, BERTopic modeling, and sophisticated classification models to improve the differentiation of misleading content. In contrast to traditional methods, our framework incorporates unique, custom linguistic features like punctuation tokens and word length, further enhancing classification precision. The model enhances topic modeling performance and content description by harnessing strict preprocessing, contextual embeddings, and dimensionality reduction. In this experiment, using Logistic Regression (LR) and eXtreme Gradient Boosting (XGBoost) classifiers, the model realized state-of-the-art performances of 92.45% accuracy and 90.80% F1-score on binary classification. The results validate the effectiveness of contextual embeddings and feature engineering in automating misinformation detection. This work gives a scalable, high-accuracy baseline for upcoming research combating misinformation and information integrity.