BCC: Bidirectional Consistency Constraint Method for Hierarchical Text Classification

Abstract

Hierarchical Text Classification (HTC) is a useful tool for document categorization based on the taxonomic hierarchy. However, current HTC methods treat labels under each category separately, which makes it difficult to model multiple inheritance labels. To solve this problem, we propose the Bidirectional Consistency Constraint (BCC) method. BCC aims to better handle multiple inheritance and class imbalance by ensuring hierarchy-compliant text-to-label mapping through relation consistency constraints and balancing loss calculation. BCC generates hierarchical features using the Multidimensional Directed Acyclic Graph (MDAG) and the Hierarchy Alignment Operator (HAO). Additionally, BCC balances the loss calculation, enabling it to fully learn the features of low-level classes (typical minor classes in HTC). Experimental results have shown that BCC outperforms state-of-the-art approaches.