Published Paper Details:

Fake news proliferation is one of the most severe issues of the digital world, affecting community opinion and destabilizing trust in media and the process of democracy [1]. Conventional machine learning and shallow word representations like TF-IDF, word2vec, and GloVe have been found to have little success in the representation of the subtle semantics and contextual relationships of news articles [2]. The recent developments in Natural Language Processing (NLP), in particular contextual embedding models like BERT, RoBERTa, XLNet, and DistilBERT, allow more comprehensive bidirectional interpretation of text, which makes them well-suited to fake news detection tasks [3], [4]. This paper performs an extensive study of the performance of these models on benchmark datasets, such as LIAR and FakeNewsNet that include both shorter political claims and longer news articles [5]. On the standard classification measures (accuracy, precision, recall, F1-score and ROC-AUC) the models are tested. Results show that RoBERTa has the best overall performance with F1-score higher than other models whereas DistilBERT has faster inference with slight accuracy compromises [6]. The results demonstrate the significance of a trade-off between accuracy and computational efficiency of applications in the real world. The paper has the following contributions: it gives a comparative framework, identifies model-specific trade-offs, and provides insights into deploying contextual embeddings in a scalable fake news detection system [7].