HeteroHTC: Enhancing Hierarchical Text Classification via Heterogeneity Encoding of Label Hierarchy

Hierarchical Text Classification (HTC) is a challenging subtask of multi-label text classification, where labels are organized into a pre-defined hierarchy. Recent works primarily encode documents and labels separately before cross attention-based feature extraction, and in the process, they collectively overlook a crucial characteristic of label hierarchies: "heterogeneity". Specifically, labels on different levels hold different granularities, and they should be projected onto distinct feature spaces; The relationships among labels are various, dictating that the message transmission among them should occur in unique feature spaces. We term these properties ''granularity heterogeneity"and "relationship heterogeneity", respectively. To fully exploit these ubiquitous yet overlooked properties, we propose HeteroHTC, which features a heterogeneous label hierarchy encoder. Additionally, we leverage pre-trained Large Language Models (LLMs) to generate high-quality label descriptions with strategically designed prompts. HeteroHTC outperforms almost all baselines in our extensive experiments on three datasets, proving its effectiveness and the necessity to take "granularity and relationship heterogeneity"into consideration.