Label-Specific Feature Augmentation for Long-Tailed Multi-Label Text Classification

Multi-label text classification (MLTC) involves tagging a document with its most relevant subset of labels from label set. In real applications, usually follow long-tailed distribution, where (called as tail-label) only contain small number documents and limit the performance MLTC. To facilitate this low-resource problem, researchers introduced simple but effective strategy, data augmentation (DA). However, existing DA approaches struggle in multi-label settings. The main reason is that augmented for one may inevitably influence other co-occurring further exaggerate problem. mitigate issue, we propose new pair-level framework MLTC, called Label-Specific Feature Augmentation (LSFA), which merely augments positive feature-label pairs tail-labels. LSFA contains two parts. first label-specific representation learning high-level latent space, second augmenting tail-label features space by transferring second-order statistics (intra-class semantic variations) head to tail labels. At last, design loss function adjusting classifiers based on datasets. whole procedure can be effectively trained. Comprehensive experiments benchmark datasets have shown proposed outperforms state-of-the-art counterparts.