search

stanfordmlgroup / chexpert-labeler

CheXpert NLP tool to extract observations from radiology reports.
MIT License
351 stars 79 forks source link

blank label #44

Open xcharxlie opened 1 year ago

xcharxlie commented 1 year ago

Thanks for your great work and datasets! I wonder how you treated label blank while training the model? As the labeler model should output only positive, negative and uncertain. Also did you use any tricks trying to fix the unbalanced dataset like data augmentation or twisted training loss functions to make model less biased?

jirvin16 commented 1 year ago

Thanks for your great work and datasets! I wonder how you treated label blank while training the model? As the labeler model should output only positive, negative and uncertain. Also did you use any tricks trying to fix the unbalanced dataset like data augmentation or twisted training loss functions to make model less biased?

Thanks for the kind words! Blank label was treated as negative (no mention of the category in the report). We did not use any methods for class balancing in our final models - the full approach is described in our paper!