Towards Robustness to Label Noise in Text Classification via Noise\n Modeling

Abstract

Large datasets in NLP suffer from noisy labels, due to erroneous automatic\nand human annotation procedures. We study the problem of text classification\nwith label noise, and aim to capture this noise through an auxiliary noise\nmodel over the classifier. We first assign a probability score to each training\nsample of having a noisy label, through a beta mixture model fitted on the\nlosses at an early epoch of training. Then, we use this score to selectively\nguide the learning of the noise model and classifier. Our empirical evaluation\non two text classification tasks shows that our approach can improve over the\nbaseline accuracy, and prevent over-fitting to the noise.\n