Based ob the paper: "UmlsBERT: Augmenting Contextual Embeddings with a Clinical Metathesaurus" (https://aclanthology.org/2021.naacl-main.139.pdf).

and the github repo: https://github.com/gmichalo/UmlsBERT

BERT base model.

Trained from scratch on MIMIC dataset, using the UMLS dataset to mask words within the text.

We achived better accuracy on MedNLI dataset.

Bert Model accuracy: 83%