This repository contains the code used for COVID-19 document type screening.

We use an XLNET model finetuned on Epistemonikos dataset consisting on non-covid evidence labeled as systematic-review, primary study non randomized controlled trial, primary study randomized controlled trial, broad synthesis and excluded. And compare to baselines BioBERT and Random Forest.

Once the model was trained we evaluate their generalization capability on CORD-19 dataset adapted to the evidence-based medicine domain.

CORD-19 dataset used for evaluation and XLNET pre-trained model with epistemonikos dataset can be found in the following link: https://zenodo.org/record/5006942

The jupyter notebook code to run models predictions are included in the following links:

• XLNET.
• BioBERT.
• Random Forest.

Jupyter notebooks for fine-tuning strategies:

• XLNET random sampling.
• XLNET data augmentation.
• XLNET uncertainty sampling iteration 1.
• XLNET uncertainty sampling iteration 2.

Results are shown in f1-score: