多粒度信息关系增强的多标签文本分类  

Multi-label Text Classification with Enhancing Multi-granularity Information Relations

作　　者：李芳芳[1] 苏朴真 段俊文 张师超 毛星亮 LI Fang-Fang;SU Pu-Zhen;DUAN Jun-Wen;ZHANG Shi-Chao;MAO Xing-Liang(School of Computer Science and Engineering,Central South University,Changsha 410038,China;Institute of Big Data and Internet Innovation,Hunan University of Technology and Business,Changsha 410205,China)

机构地区：[1]中南大学计算机学院,湖南长沙410038 [2]湖南工商大学大数据与互联网创新研究院,湖南长沙410205

出　　处：《软件学报》2023年第12期5686-5703,共18页Journal of Software

基　　金：国家自然科学基金(62172449,61836016,71790615,62006251,62172441);湖南省自然科学基金(2021JJ30870,2021JJ40783);长沙市自然科学基金(kq2014134);国防科技重点实验室基金(6142101190302)。

摘　　要：基于深度学习的多标签文本分类方法存在两个主要缺陷:缺乏对文本信息多粒度的学习,以及对标签间约束性关系的利用.针对这些问题,提出一种多粒度信息关系增强的多标签文本分类方法.首先,通过联合嵌入的方式将文本与标签嵌入到同一空间,并利用BERT预训练模型获得文本和标签的隐向量特征表示.然后,构建3个多粒度信息关系增强模块:文档级信息浅层标签注意力分类模块、词级信息深层标签注意力分类模块和标签约束性关系匹配辅助模块.其中,前两个模块针对共享特征表示进行多粒度学习:文档级文本信息与标签信息浅层交互学习,以及词级文本信息与标签信息深层交互学习.辅助模块通过学习标签间关系来提升分类性能.最后,所提方法在3个代表性数据集上,与当前主流的多标签文本分类算法进行了比较.结果表明,在主要指标Micro-F1、Macro-F1、nDCG@k、P@k上均达到了最佳效果.Multi-label text classification methods based on deep learning lack multi-granularity learning of text information and the utilization of constraint relations between labels.To solve these problems,this study proposes a multi-label text classification method with enhancing multi-granularity information relations.First,this method embeds text and labels in the same space by joint embedding and employs the BERT pre-trained model to obtain the implicit vector feature representation of text and labels.Then,three multi-granularity information relations enhancing modules including document-level information shallow label attention(DISLA)classification module,word-level information deep label attention(WIDLA)classification module,and label constraint relation matching auxiliary module are constructed.The first two modules carry out multi-granularity learning from shared feature representation:the shallow interactive learning between document-level text information and label information,and the deep interactive learning between word-level text information and label information.The auxiliary module improves the classification performance by learning the relation between labels.Finally,the comparison with current mainstream multi-label text classification algorithms on three representative datasets shows that the proposed method achieves the best performance on main indicators of Micro-F1,Macro-F1,nDCG@k,and P@k.

关 键 词：注意力机制 多标签文本分类 标签关系 多粒度信息 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]