Post-2019 MeSH Terms
I'm a data scientist working on a project for the NLM to automatically assign MeSH terms to catalog records using WellcomeBertMesh.
The model works well overall but since it was trained on data up to 2019, it has no knowledge of MeSH terms added after that. This is a significant gap for our use case as many of our records cover these newer topics.
Is there a way to support post-2019 MeSH terms without fully retraining the model? like COVID-19
For example, fine-tuning on new terms or updating the label vocabulary in a meaningful way?
Any guidance would be appreciated. Thank you!
Hi Irtazaaslam,
Exciting to hear NLM is using our model, unfortunately we haven't yet retrained the MeSH model on more up to date MeSH terms.
My only suggestion would be to either fine-tune only the model classifier head to predict on these updated classes.
We also have a more simple linear model that gives better results across the classes.
Happy to discuss further,
Many thanks,
Justin Boylan-Toomey
Lead Machine Learning Engineer
Wellcome Trust