Predicting COVID-19 Related Tweets Using Ensemble of Transformers Models

In the past few years, COVID-19 has been consid-ered one of the most dangerous pandemics in several countries. There is a lot of information circulating on social media platforms about COVID-19, some of it is reliable, while others may be exag-gerated or unfounded. Using machine learning-driven sentiment analysis is considered a valuable tool that helps understand the community's feelings regarding many issues like the COVID-19 outbreak. Developing an accurate model that can assess if a tweet is about COVID-19 is a challenging task. This study aims to classify the tweets whether it is about COVID-19 or not using deep learning and transformers models. The developed model improves the gathering of tweets data about the COVID-19 epidemic without relying only on keywords such as ‘covid’ or ‘coronavirus'. In this work, we proposed the best model based on an ensemble method that effectively combines three models which are: BERTweet-covid19-base-cased, BERTweet, and RoBERTa. We applied the models to the data set provided by the Zindi community. The best results were achieved over the tested dataset in terms of Log-Loss with a minimum value of 0.154,0.174,0.170, and 0.191 for the proposed ensemble model, BERTweet-covid19-base-cased, BERTweet, and RoBERTa respectively. Our proposed model is ranked first among all the participant teams.