Publication

Abstract :

Social media is arguably the single largest and most dynamic generator of information that is directly and immediately ingested into the psyche of the society. The peculiar nature of social media information is its immediate and blind consumption without any attempt at validating its veracity or the reliability of its provenance. With social media platforms expanding their support to regional languages, the evil of fake information has acquired larger dimensions. Fake news is a major source of misinformation, challenging in itself as a computational problem and more so in regional languages. This is because of the vast cultural, structural and semantic differences between languages, making state of art models are unusable or requiring significant retraining and tweaking. In this paper, a novel hybrid CNN-BiLSTM model using FastText word embedding along with lime explainability approach for rendering transparency is proposed for fake news detection in Malayalam language, a dravidian language spoken in state of Kerala. It gave an F1 Score of 0.90, outperforming existing models. Also the performance was analysed using other deep learning techniques incorporating attention mechanism. The work can be extended to other regional languages in future.