Sentence-level Direct Assessment QE Shared Task 2020 Forum

Go back to competition Back to thread list Post in this thread

> About z-standardised DA scores


About the following sentence:
"We thus expect an absolute quality score for each sentence translation (z-standardised DA)."

Are we suppposed to z-standardize DA scores before submission as if the predictor is a human annotator?

I built models predicting z-scores on the training set directly, which did not achieve high scores.
I then built models predicting DA scores, which achieve MAE of ~13 while the leader board contains entries having MAE of 0.5.

Are we expected to predict DA-scores and then z-standardize them before submission via (x - x.mean()) / x.std() ?

Posted by: bicici @ July 16, 2020, 3:16 p.m.

Hi Ergun,
according to your submissions, your sentence indices start at 1,
while they should start at 0. This would explain why you get those
scores.

I sent you this information by email yesterday already.

Best,
Fred

Posted by: fblain @ July 17, 2020, 2 p.m.
Post in this thread