Data leakage and duplication
Thank you for your work, I like it a lot! But it has a data leakage issue.
In UD treebanks there is a consistent issue of data duplication within and across splits, leading to data leakage from train into test splits. This is also the case in this dataset unfortunately. After deduplicating and checking overlap between all splits for en_ewt I see:
Removed 809 samples from training set! (12,543->11,734)
Removed 123 samples from validation set! (2,001->1,878)
Removed 165 samples from test set! (2,077->1,912)
Train & validation overlap: 35
Train & test overlap: 46
Validation & test overlap: 27
Since the goal of this project is making available golden datasets for training models, I highly encourage a thorough deduplication of the data so that we have actually usable datasets that are provided in a canonical manner. Otherwise some people will dedupe (each in their own way), others will not, and ultimately results will not be comparable.