BBSN: Bilateral-Branch Siamese Network for Imbalanced Multi-label Text Classification

Abstract

In multi-label text classification, the numbers of instances in different categories are usually extremely imbalanced. How to learn good models from imbalanced data is a challenging task. Some existing works tackle it through class re-balancing strategies or imbalanced loss objectives, but their performance remains limited in the cases of imbalanced distributed data. In this work, we propose a model, which combined Siamese Network and Bilateral-Branch Network to deal with both representation learning and classifier learning simultaneously. In the siamese network component, we propose a category-specific similarity strategy to improve the representation learning and adapt a novelty dynamic learning mechanism to make the model end-to-end trainable, and in the bilateral-branch network, we adopt the cumulative learning strategy to shift the learning focus from universal pattern to tail learning. In general, we adopt a multi-task architecture to ensure that both the head categories and the tail categories are adequately trained. The experiments on two benchmark datasets show that our method can improve the performance on the entire and tail categories, and achieves competitive performance compared with existing approaches.