基于多尺度特征提取的层次多标签文本分类方法  

Hierarchical Multi-label Text Classification Method Based on Multi-scale Feature Extraction

作　　者：武子轩 王烨 于洪[1] WU Zixuan;WANG Ye;YU Hong(Chongqing Key Laboratory of Computational Intelligence,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学计算智能重庆市重点实验室,重庆400065

出　　处：《郑州大学学报（理学版）》2025年第2期24-30,共7页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家重点研发计划(2021YFF0704100);国家自然科学基金项目(62136002,62233018);重庆市自然科学基金项目(cstc2022 ycjh-bgzxm0004)。

摘　　要：针对现有的特征提取方法忽略文本局部和全局联系的问题,提出了基于多尺度特征提取的层次多标签文本分类方法。首先,设计了多尺度特征提取模块,对不同尺度特征进行捕捉,更好地表示文本语义。其次,将层次特征嵌入文本表示中,得到具有标签特征的文本语义表示。最后,在标签层次结构的指导下对输入文本构建正负样本,进行对比学习,提高分类效果。在WOS、RCV1-V2、NYT和AAPD数据集上进行对比实验,结果表明,所提模型在评价指标上表现出色,超过了其他主流模型。此外,针对层次分类提出层次Micro-F 1和层次Macro-F 1指标,并对模型效果进行了评价。A hierarchical multi-label text classification method based on multi-scale feature extraction was proposed to address the issue of current feature extraction methods in neglecting the local and global connections in text.Firstly,a multi-scale feature extraction module was designed to capture features at different scales,aiming to provide a better representation of text semantics.Secondly,the hierarchical features were embedded into the text representation to obtain a text semantic representation with label features.Finally,with the guidance of the label hierarchy,positive and negative samples were constructed for the input text,and contrastive learning was performed to enhance the classification effectiveness.Comparative experiments were conducted on the WOS,RCV1-V2,NYT and AAPD datasets.The results indicated that the proposed model performed well in terms of the evaluation indices and exceeded other mainstream models.Additionally,the hierarchical Micro-F 1 and Macro-F 1 indicators were proposed for hierarchical classification,and the effectiveness of the model was evaluated.

关 键 词：层次多标签文本分类 多尺度特征提取 对比学习 层次Micro-F 1 层次Macro-F 1 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]