HiMTL: Hierarchy-aware Multi-Task Learning for Hierarchical Text Classification

ACL ARR 2025 February Submission2381 Authors

Abstract: Hierarchical relationships between labels can be used to control the information flow in text classification models. However, while these models are required to distinguish nuances between closely related fine-grained labels, their weight updates are also influenced by unrelated branches of the hierarchy. In this paper, we show that systematically splitting the hierarchy into multiple sub-hierarchies, thus training multiple localized hierarchy-aware classification layers on top of a shared text encoder, can improve classification scores on simple and complex hierarchies. HiMTL is not a new model but an architectural extension that can be applied to different state-of-the-art hierarchical classification models.

Paper Type: Short

Research Area: Machine Learning for NLP

Research Area Keywords: graph-based methods, multi-task learning, structured prediction

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 2381