Performance on trec dataset drops significantly after using eda about eda_nlp

[jasonwei20]

If you have manually checked several examples of augmented data and their labels and those are correct, then your code for generating the augmented data is probably correct.

I am not sure about how EDA affects the performance of BERT. One reason I could think of that might hurt the performance is if BERT already has the ability to do well on TREC and SST (which it probably does), i.e., there is little or no overfitting on BERT, then introducing some augmentations might actually hurt its performance. To be conservative, you can try using a really low alpha parameter (0.05 or 0.03).

Sorry I can't be more helpful on this occasion. If you find that EDA does not do well versus long or short texts, and want to report some results on that, I'm happy to modify the readme so that others know.

[RenShuhuai-Andy]

Thanks very much for your reply. I will do more experiments based on BERT and give feedback if there are new discoveries.