uda icon indicating copy to clipboard operation
uda copied to clipboard

Published 20 hours ago • verified publisher icon google-research

Results on RTE/SNLI/MNLI tasks

Open 1024er opened this issue 4 years ago • 2 comments

I tried the code on RTE/SNLI/MNLI tasks, but UDA's results are worse, did anyone tried it before ?

1024er avatar Jul 28 '19 06:07 1024er

@qizhex will you please help me ? Thank you

1024er avatar Jul 30 '19 08:07 1024er

Hi, I haven't tried it before because there are no unlabeled data for it. Did you use part of the training set as the unlabeled data? If I were you, I would first verify that back translation works for supervised learning for these tasks. If it helps for supervised learning, then it's likely that it works in semi-supervised learning setting. Note that you might need to use a very low temperature because these tasks might not be noise tolerant.

michaelpulsewidth avatar Jul 30 '19 20:07 michaelpulsewidth