Weakly supervised text classification framework for noisy-labeled imbalanced samples

The goal of this study is to solve the combined issue of noise labels and imbalanced samples for text classification. Current studies generally adopt data sampling or cleaning in model learning, leading to a part of information loss. To this end, this paper introduces a weakly supervised text classification framework, dubbed WeStcoin, which aims to learn a clear hierarchical attention network directly from the given noisy-labeled imbalanced samples. Specifically, WeStcoin first vectorizes the given texts to generate a contextualized corpus on which the pseudo-label vector is calculated by extracting seed words from each class and the predicted label vector is obtained by a hierarchical attention network. Based on the pseudo and predicted label vectors, we learn a cost-sensitive matrix to project the concatenated label vectors into the given label space. WeStcoin is trained iteratively to reduce the difference between the output labels and the given noisy labels by updating the network parameters, the set of seed words, and the cost-sensitive matrix, respectively. Finally, extended experiments on short-text classification shows that WeStcoin achieves a significant improvement than the state-of-the-art models in imbalanced samples with noisy labels. Besides, WeStcoin acts more robustly than compared methods and provides potential explanations for noisy labels.