tldr; #data #augmentation in #NLProc degrades #textClassification performance in most cases. https://aclanthology.org/2022.insights-1.12.pdf

Well that was fun. Just spent last night experimenting with #dataAugmentation (using #flan-T5 for paraphrasing) and in wondering why it seems to degrade #textClassification performance I came across this great paper essentially saying the same thing. I guess I’ll revisit this in a year or so when there are better language models.
@linguistics @sigmoid.social