UDA for Sequence Classification problems (like NER)

[kalpitdixit]

Hi,

Regarding UDA for NLP-SequenceTagging like NER. Any ideas for what kind of augmentation methods would work for them?

e.g. BackTranslation is tough to use for NER because,

1. we would need to find some way to project labels from the original tokens to the augmented tokens (which might even be different in number)
2. in general, for each augmented token, its not clear which original token it should be compared against (especially since the number of tokens itself might change)

Thanks.

[guotong1988]

Same question.

[xxllp]

+1