Scholarlink

<p>Hierarchical Text Classification (HTC) can be applied in various domains, such as news topic classification and academic paper classification. The traditional approach treats HTC as a multi-label classification problem, ignoring the hierarchical structure of the labels. The latest HTC approaches inject the label hierarchy into the text encoder by contrastive learning, making the text features and label features closer. However, the generation of samples in contrastive learning introduces noise and overlooks the correlations between labels within the same batch. To overcome this limit, we propose Dual Contrastive Learning (DCL) for HTC, which combines hierarchy-guided and label-enhanced contrastive learning. During the training in DCL, hierarchy-guided contrastive learning pulls together the input text and positive samples composed of keywords, enabling the text encoder to learn hierarchical-aware text representations. Meanwhile, label-enhanced contrastive learning pulls together the embeddings of similar labels within a batch, fully utilizing the correlations between labels. To demonstrate the efficacy of our proposed model, we perform extensive experiments on three publicly available datasets and one real-world dataset. Our code is available at <a href="https://github.com/man-123456/Dual-HTC-main.git">https://github.com/man-123456/Dual-HTC-main.git</a>.</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

Dual contrastive learning for hierarchical text classification

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Qiang Xu,
  • Jiaxiang Man,
  • Jizhong Xi,
  • Shengwei Ji

摘要

Hierarchical Text Classification (HTC) can be applied in various domains, such as news topic classification and academic paper classification. The traditional approach treats HTC as a multi-label classification problem, ignoring the hierarchical structure of the labels. The latest HTC approaches inject the label hierarchy into the text encoder by contrastive learning, making the text features and label features closer. However, the generation of samples in contrastive learning introduces noise and overlooks the correlations between labels within the same batch. To overcome this limit, we propose Dual Contrastive Learning (DCL) for HTC, which combines hierarchy-guided and label-enhanced contrastive learning. During the training in DCL, hierarchy-guided contrastive learning pulls together the input text and positive samples composed of keywords, enabling the text encoder to learn hierarchical-aware text representations. Meanwhile, label-enhanced contrastive learning pulls together the embeddings of similar labels within a batch, fully utilizing the correlations between labels. To demonstrate the efficacy of our proposed model, we perform extensive experiments on three publicly available datasets and one real-world dataset. Our code is available at https://github.com/man-123456/Dual-HTC-main.git.