HiSLR: Sample-Specific Relationships in Hyperbolic Space for Hierarchical Text Classification

Existing approaches to Hierarchical Text Classification (HTC) often rely on encoding the global hierarchical structure to generate label features, adding unnecessary complexity since the entire hierarchy is not always relevant to each sample. Instead, the hierarchical structure of labels specific to each text sample is more critical, as it captures relevant relationships while disregarding unrelated parts of the global hierarchy. These relationships can be modeled as the semantic alignment of the text with its positive labels, organized hierarchically. However, without explicitly encoding the global hierarchy, achieving this alignment directly in Euclidean space is challenging, as its flat geometry does not naturally support hierarchical relationships. To address this, we propose Hierarchical Sample-specific Label Relationships (HiSLR), which utilizes the Lorentz model to learn the implicit sample-specific hierarchy. HiSLR projects text and label features into hyperbolic space using an exponential map transformation. It applies contrastive loss to ensure semantic alignment between the text and its positive labels. This alignment, based on distances in hyperbolic space, inherently captures hierarchical relationships and eliminates the need for explicit global hierarchy encoding. Experimental results on four benchmark datasets validate the superior performance of HiSLR over baseline methods.