Multi-Label Text Classification with Large Language Models via Hierarchical Prompt Learning and Label Dependency Modeling

Abstract

This study proposes a large language model enhancement method based on hierarchical prompt learning to address the challenges of multi-label text classification, including strong label dependencies, complex hierarchical relations, and data imbalance. The study first analyzes the limitations of traditional approaches in flat classification and label correlation modeling, pointing out their weaknesses in handling complex semantics and fine-grained labels. To address this problem, a hierarchical prompt mechanism is designed. The label system is mapped into a multi-level prompt structure, which guides the model step by step from macro categories to fine-grained labels, enabling better capture of both global and local semantic features. The method introduces an interaction mechanism between prompts and text representations within the model structure. Through attention-based fusion, the model dynamically learns different levels of semantic information. A hierarchical classifier is then applied to complete multi-level label prediction. The experimental part includes multiple comparative and sensitivity experiments. These cover hyperparameter settings, regularization strength, noise label injection rates, and class imbalance ratios. The results show that the proposed method significantly outperforms existing approaches in AUC, ACC, F1-Score, and Precision. It also maintains strong robustness under various complex conditions. The study demonstrates the effectiveness of combining hierarchical prompt learning with large language models. It not only improves the overall performance of multi-label classification but also shows stability and reliability across different experimental setups. This provides a feasible, enhanced path for text processing in complex semantic environments.