Can NLP augmentation affect the performance of certain categories of data?

There are some few different kinds of subcategories in my NLP data.
My NLP Model performance is not that high because some subcategories.
I expect to increasing model performance via increasing subcategories data by NLP augmentation. Does it make sense?
Just worried about thing is even if distribution of subcategories hasn’t unbalanced, NLP augmentation could be helpful?
Is Increasing the subcategories data good solution ignoring whole data balance?

Data augmentation can be a suitable way to address imbalanced datasets. However, the question is: How exactly would you do this? I would argue that this for from trivial form NLP-related task.

In case of images, you can: rotate, crop, squeeze, change color depth, decrease/increase hue, etc. Assuming your inputs are sentences, how would you “tweak” this sentences to get additional and meaningful input sentences?

Thank you for repling.
I’m trying to apply ContextualWordEmbsAug(word insert or substitute) as word augmenter based on BERT using nlpaug

Thanks for the link. Looking at the example from their Github page kind of confirms my concerns:

Data augmentation for sentences is not easy when the goal is the get semantically meaningful outputs. Even reasonably straightforward synonym replacement is not foolproof. In the table above, 2nd row, it works well enough. But for example, let’s say the original sentence is “the horse jumps over the fence” and the augmented output is “the horse climbs over the fence”. It doesn’t really seem semantically meaningful. Not sure if I would want to train a language model based on this.

That being said, those are more my own gut feelings. I assume the authors have quantitative result to show the benefits of their approaches.