Readers versus Re-rankers in Question Answering over COVID-19 scientific literature

Resumen

In this work we present a comparison between the two most used neural Question Answering (QA) architectures to solve the problem of information overload on COVID-19 related articles. The span extraction (reader) and the re-ranker. We have found that there are no studies that compare these two methods even though they are so widely used. We also performed a search of the best hyperparameters for this task, and tried to conclude whether a model pre-trained with biomedical documents such as bioBERT outperforms a general domain model such as BERT.We found that the domain model is not clearly superior to the generalist one. We have studied also the number of answers to be extracted per context to obtain consistently good results. Finally, we conclude that although both approaches (readers and rerankers) are very competitive, readers obtain systematically better results.