Classification of COVID-19 medical articles using deep learning model

Abstract

The global pandemic of Corona Virus Disease 19 (COVID-19) has made an impact on our daily life. After 2019, the literatures that focus on COVID-19 have rising exponentially. It is almost impossible for human to read all literatures and classify them. In this article, we propose the method to make an unsupervised model called zero-shot classification model from pre-trained BERT (Bidirectional Transfomers) model. We use CORD-19 dataset in conjunction with LitCovid database for construct new vocabulary and prepare test dataset. For Natural Language Inference (NLI) downstream task, we use three corpus – Standford Natural Language Inference (SNLI), Multi-Genre Natural Language Inference (MultiNLI) and MedNLI. We can significantly reduce the training time to build a task specific machine learning model by 98.2639%. The final model can run faster and use lower resources than the comparators. It has 27.84% accuracy which is lower than the best achieve accuracy by 6.73%, but it is comparable. Finally, we can identify that tokenizer and vocabulary that is more specific to COVID-19 do not outperform the generalization one, also BART architecture affects the classification result too.