How do you get around lack of data for text categorization/NER data? Is this a painpoint for you?

For a lot of clients, I see issues with some categories of data being severely underrepresented in the training data.

For instance, a category could just have 10 examples in a total dataset size of 10,000. In such, cases do you downsample, generate synthetic examples etc?

What would be the ideal solution?

data datascience ner text