Mechanism to choose between EDA tasks about eda_nlp

[lolo8304]

I also agree. I have implemented eg that ‘remove’ only works if number of words in sentence > 6 because it would change the meaning.

Insertion of synonyms at another place is not an issue.
I have then used different algo predictions if 3 or 4 algo are used.

I am using the EDA algorithm to create 7 new sentences to extend the nun er of utterances per intent to train with LUIS / Microsoft.
With great success.

Overfitting or overlapping intents:
If 2 intents are too close together we anyway propose the customer to choose which one one prefers.

[jasonwei20]

Ah, you're right, COLA is a good example of a dataset where EDA might not the best type of augmentation to use.

I'm not actively maintaining the repository right now but if you'd like to send a PR, I can certainly do a code review.