Self-Refine Learning For Data-Centric Deep Learning

In industry NLP application, our manually labeled data has a certain number of noise data. We present a simple method to find the noise data and remove them. We select the noise data whose human label is not contained in the top-K model’s predictions. The experiment result shows that our method works. For industry deep learning application, our method improve the text classification accuracy from 80.5% to 90.6% in dev dataset, and improve the human-evaluation accuracy from 83.2% to 90.1%. The conclusion is: The self-predict and remove/reset method of this paper can not improve the accuracy to more than 95% of human evaluation, without human labeling again for the training dataset.