ABSTRACT
In view of COVID-19 outbreak, the world is facing lot of issues related to public health. Online media and platforms especially during the present pandemic have increased the popularity of many online applications and also blogs. Few people are using this opportunity for the good cause, whereas few others are misusing social media to share fake news and false information about the pandemic. The main idea behind sharing fake news may be to mislead communities, individuals, countries, etc. for various reasons like political, economic, or even for fun. Such fake news and false information impact the society negatively and can cause distrust in public. Detecting fake news and avoiding the spread of the same in social media is posing a big challenge. Even though researchers have explored several tools and techniques to address fake news and hostile posts in various domains, it is still an open problem as there will always be a new domain like COVID-19. In view of this, this paper describes two models based on transfer learning (TL) approaches, namely extended universal language model fine-tuning (Ext-ULMFiT) and fine-tuned bidirectional encoder representations from transformers (FiT-BERT). Both the models are fine-tuned on CORD-19 dataset to combat COVID-19 fake news. The proposed models evaluated on COVID-19 fake news detection shared task dataset of CONSTRAINT’21 workshop obtained 0.99 weighted average F1 score. However, FiT-BERT outperformed Ext-ULMFiT in predicting fake news’ and Ext-ULMFiT was more successful in the prediction of real news. Further, the performances of the proposed models are very close to the best performing team of COVID-19 fake news detection shared task in CONSTRAINT’21 workshop. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.