ATIS data split (train,dev,test) issues
kpe opened this issue · comments
I noticed some issues with the data split in the ATIS dataset (see visualization of the label distributions here.):
- duplicated data samples - 397 (from 5871)
- no train data for 5 intent and 7 slot labels
- up to 20% of the labels present in the training dataset are not present at all in the dev or test datasets
I believe this could make the interpretation of model performance measures somewhat unreliable, and tried to build an alternative, more balanced
data split here.