Affiliations:
Department of Information Technology, Sri Sivasubramaniya Nadar College of Engineering, Kalavakkam, India
Correspondence:
[*]
Corresponding author. M. Badri Narayanan, Department of Information Technology, Sri Sivasubramaniya Nadar College of Engineering, Kalavakkam, 603110, India. E-mail: badrinarayananm18018@it.ssn.edu.in
Abstract: Fake news production, accessibility, and consumption have all increased with the rise of internet-connected gadgets and social media platforms. A good fake news detection system is essential because the news readers receive can affect their opinions. Several works on fake news detection have been done using machine learning and deep learning approaches. Recently, the deep learning approach has been preferred over machine learning because of its ability to comprehend the intricacies of textual data. The introduction of transformer architecture changed the NLP paradigm and distinguished itself from recurrent models by enabling the processing of sentences as a whole rather than word by word. The attention mechanisms introduced in Transformers allowed them to understand the relationship between far-apart tokens in a sentence. Numerous deep learning works on fake news detection have been published by focusing on different features to determine the authenticity of a news source. We performed an extensive analysis of the comprehensive NELA-GT 2020 dataset, which revealed that the title and content of a news source contain discernible information critical for determining its integrity. To this objective, we introduce ‘FakeNews Transformer’ — a specialized Transformer-based architecture that considers the news story’s title and content to assess its veracity. Our proposed work achieved an accuracy of 74.0% on a subset of the NELA-GT 2020 dataset. To our knowledge, FakeNews Transformer is the first published work that considers both title and content for evaluating a news article; thus, we compare the performance of our work against two BERT and two LSTM models working independently on title and content. Our work outperformed the BERT and LSTM models working independently on title by 7.6% and 9.6%, while performing better than the BERT and LSTM models working independently on content by 8.9% and 10.5%, respectively.
