SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification

Hierarchical text classification (HTC) is a challenging task that requires classifiers to solve a series of multi-label subtasks considering hierarchical dependencies among labels. Recent studies have introduced prompt tuning to create closer connections between the language model (LM) and the compl...

Full description

Saved in:
Bibliographic Details
Main Authors: He Zhu, Jinxiang Xia, Ruomei Liu, Bowen Deng
Format: Article
Language:English
Published: MDPI AG 2025-01-01
Series:Entropy
Subjects:
Online Access:https://www.mdpi.com/1099-4300/27/2/128
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850081616218554368
author He Zhu
Jinxiang Xia
Ruomei Liu
Bowen Deng
author_facet He Zhu
Jinxiang Xia
Ruomei Liu
Bowen Deng
author_sort He Zhu
collection DOAJ
description Hierarchical text classification (HTC) is a challenging task that requires classifiers to solve a series of multi-label subtasks considering hierarchical dependencies among labels. Recent studies have introduced prompt tuning to create closer connections between the language model (LM) and the complex label hierarchy. However, we find that the model’s attention to the prompt gradually decreases as the prompt moves from the input to the output layer, revealing the limitations of previous prompt tuning methods for HTC. Given the success of prefix tuning-based studies in natural language understanding tasks, we introduce <b>S</b>tructural entro<b>P</b>y gu<b>I</b>ded p<b>R</b>ef<b>I</b>x <b>T</b>uning (<b>SPIRIT</b>). Specifically, we extract the essential structure of the label hierarchy via structural entropy minimization and decode the abstractive structural information as the prefix to prompt all intermediate layers in the LM. Additionally, a depth-wise reparameterization strategy is developed to enhance optimization and propagate the prefix throughout the LM layers. Extensive evaluation on four popular datasets demonstrates that SPIRIT achieves a state-of-the-art performance.
format Article
id doaj-art-8f1b0df5edea48a4b3dc4fa38e4cd39f
institution DOAJ
issn 1099-4300
language English
publishDate 2025-01-01
publisher MDPI AG
record_format Article
series Entropy
spelling doaj-art-8f1b0df5edea48a4b3dc4fa38e4cd39f2025-08-20T02:44:42ZengMDPI AGEntropy1099-43002025-01-0127212810.3390/e27020128SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text ClassificationHe Zhu0Jinxiang Xia1Ruomei Liu2Bowen Deng3State Key Laboratory of Software Development Environment, School of Computer Science and Engineering, Beihang University, No. 37 Xue Yuan Road, Hai Dian District, Beijing 100191, ChinaState Key Laboratory of Software Development Environment, School of Computer Science and Engineering, Beihang University, No. 37 Xue Yuan Road, Hai Dian District, Beijing 100191, ChinaState Key Laboratory of Software Development Environment, School of Computer Science and Engineering, Beihang University, No. 37 Xue Yuan Road, Hai Dian District, Beijing 100191, ChinaSchool of Advanced Technology, Xi’an Jiaotong-Liverpool University, Suzhou 215213, ChinaHierarchical text classification (HTC) is a challenging task that requires classifiers to solve a series of multi-label subtasks considering hierarchical dependencies among labels. Recent studies have introduced prompt tuning to create closer connections between the language model (LM) and the complex label hierarchy. However, we find that the model’s attention to the prompt gradually decreases as the prompt moves from the input to the output layer, revealing the limitations of previous prompt tuning methods for HTC. Given the success of prefix tuning-based studies in natural language understanding tasks, we introduce <b>S</b>tructural entro<b>P</b>y gu<b>I</b>ded p<b>R</b>ef<b>I</b>x <b>T</b>uning (<b>SPIRIT</b>). Specifically, we extract the essential structure of the label hierarchy via structural entropy minimization and decode the abstractive structural information as the prefix to prompt all intermediate layers in the LM. Additionally, a depth-wise reparameterization strategy is developed to enhance optimization and propagate the prefix throughout the LM layers. Extensive evaluation on four popular datasets demonstrates that SPIRIT achieves a state-of-the-art performance.https://www.mdpi.com/1099-4300/27/2/128information theorystructural entropyrepresentation learning
spellingShingle He Zhu
Jinxiang Xia
Ruomei Liu
Bowen Deng
SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification
Entropy
information theory
structural entropy
representation learning
title SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification
title_full SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification
title_fullStr SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification
title_full_unstemmed SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification
title_short SPIRIT: Structural Entropy Guided Prefix Tuning for Hierarchical Text Classification
title_sort spirit structural entropy guided prefix tuning for hierarchical text classification
topic information theory
structural entropy
representation learning
url https://www.mdpi.com/1099-4300/27/2/128
work_keys_str_mv AT hezhu spiritstructuralentropyguidedprefixtuningforhierarchicaltextclassification
AT jinxiangxia spiritstructuralentropyguidedprefixtuningforhierarchicaltextclassification
AT ruomeiliu spiritstructuralentropyguidedprefixtuningforhierarchicaltextclassification
AT bowendeng spiritstructuralentropyguidedprefixtuningforhierarchicaltextclassification