UDA(https://github.com/google-research/uda) could achieve good accuracy by only 20 training data on text classification.
But I find it is hard to reproduce the result on my own dataset.
So I want to know the reason why UDA works. And I want to know what is the most important thing to reproduce the result.