Data Augmentation can boost the performance of an intent classifier. In this post, we will introduce a data augmentation method.
This method is proposed in paper: New Intent Discovery with Pre-training and Contrastive Learning
How to implement data augmentation in intent classification?
This method is simple, it is:
As to a sentence, we can use some words to replace wods in this sentence randomly.
However, there is a problem, how many words should be replaced in a sentence?
From this paper, we can find 25% words that are replaced will make the model get a good performance.