Scholarlink

<p>Multilabel text classification (MLTC) requires capturing the relationships between texts and labels. Unlike single-label tasks, it has a higher label co-occurrence rate. We propose GSBC-SNet, an end-to-end graph convolutional network. It fuses the label information of GCN and the features of RoBERTa through a unique attention mechanism, breaking the traditional fusion approach. Its adaptive adjacency matrix, designed by a novel algorithm, focuses on the actual label correlations in the text rather than just co-occurrence. This matrix can reduce noise and alleviate the long-tail effect in label distribution, adapting to diverse data. The positive–negative sample mechanism optimizes the learning process by leveraging the synergy among labels, context, and the matrix. Experiments on the AAPD, RCV1, and Reuters-21578 datasets demonstrate that GSBC-SNet outperforms existing methods and achieves better scores in some metrics on the AAPD dataset. This research enhances the model performance and provides new insights for handling label correlations, thus promoting the development of the MLTC field.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GSBC-SNet: a novel graph-aware bidirectional contrastive semantic network for multilabel text classification

  • 期刊论文

  • 发表日期：

    2025年4月29日

查看全文
  • Rui Pang,
  • Qiongbing Zhang,
  • Yating Lin,
  • Liang Ouyang,
  • Zhangwei Cui

摘要

Multilabel text classification (MLTC) requires capturing the relationships between texts and labels. Unlike single-label tasks, it has a higher label co-occurrence rate. We propose GSBC-SNet, an end-to-end graph convolutional network. It fuses the label information of GCN and the features of RoBERTa through a unique attention mechanism, breaking the traditional fusion approach. Its adaptive adjacency matrix, designed by a novel algorithm, focuses on the actual label correlations in the text rather than just co-occurrence. This matrix can reduce noise and alleviate the long-tail effect in label distribution, adapting to diverse data. The positive–negative sample mechanism optimizes the learning process by leveraging the synergy among labels, context, and the matrix. Experiments on the AAPD, RCV1, and Reuters-21578 datasets demonstrate that GSBC-SNet outperforms existing methods and achieves better scores in some metrics on the AAPD dataset. This research enhances the model performance and provides new insights for handling label correlations, thus promoting the development of the MLTC field.