CoViNAR: A Context-Aware Social Media Dataset for Pandemic Severity Level Prediction and Analysis

Soofi Shafiya^1*Mudasir Ahmad Wani²Suraiya Jabin¹Mohammed ELAffendi²J Jahiruddin¹

• ¹Jamia Millia Islamia, Delhi, India
• ²Prince sultan university, riyadh, Saudi Arabia

The unprecedented COVID-19 pandemic exposed critical weaknesses in global health management, particularly in resource allocation and demand forecasting. This work introduces a transformative approach to enhancing pandemic preparedness through real-time social media analysis. Using SnScrape, over 27.5 million tweets for the duration of November 2019 to March 2023 were collected using COVID-19-related hashtags. Tweets from April 2021, a peak pandemic period, were selected to create the CoViNAR dataset. BERTopic enabled context-aware filtering, resulting in a novel dataset of 14,000 annotated tweets categorized as "Need", "Availability", and "Not-relevant." The CoViNAR dataset was used to train various machine learning classifiers, and the best classifier achieved an accuracy of 96.42%, 96.44% precision, 96.42% recall, and an F1-score of 96.43% on the Test dataset. While training the NAR classifier, we experimented with three context-aware word embedding techniques, with DistilBERT yielding the best performance. We demonstrated the success of the NAR classifier by performing a temporal analysis of tweets from the US, UK, and India from November 2019 to March 2023. The strong correlation between NAR tweet counts and COVID-19 case surges highlighted the potential of the proposed method, offering health authorities a powerful, proactive tool for resource management during a pandemic.