Label Hierarchy Alignment for Improved Hierarchical Text Classification

Ashish Kumar, Durga Toshniwal

Abstract: Effective text and label representations are crucial for accurate predictions in Hierarchical Text Classification (HTC). However, existing methods face challenges in capturing relevant label representations due to the discrete nature of labels and the complexities of their hierarchical relationships. To address these challenges, we propose Label Hierarchy Alignment (LHA), which ensures that learned label representations conform to the hierarchical label space. In this paper, we present two LHA approaches. The first is adversarial label alignment, which utilizes adversarial learning to enforce a prior distribution on label embeddings. This guides the model to adjust label embeddings in accordance with the hierarchical structure. The second approach is contrastive label alignment, which uses contrastive learning to identify semantic similarities and differences between label nodes across the hierarchy. This enriches the label embeddings with semantic depth, thereby enhancing their alignment with the hierarchical structure. By explicitly addressing various hierarchical aspects, both approaches generate refined label representations that align more effectively with the label hierarchy. We implemented LHA on the current state-of-the-art model and evaluated its performance on two benchmark datasets. The experiment results show that performing LHA improves HTC performance.