Scholarlink

<p>Multi-label text classification underpins applications such as information retrieval, news tagging, and scholarly annotation, yet existing methods remain vulnerable to noise and redundancy, tend to model label-–text and text-–text relations from a single view, and insufficiently exploit document hierarchies. We propose LTC-MPE, a dual-branch co-attention model with evolving label prototypes. A pretrained encoder first produces contextual token representations, upon which a label-guided word filter constructs a binary mask that sparsifies sequences, suppresses noise. On this basis, two parallel branches derive a label-specific view via label attention and a sentence-level hierarchical view capturing local and global dependencies. A bidirectional co-attention module then deeply aligns and adaptively fuses these complementary representations. Finally, label prototypes are updated online via a momentum mechanism and guide document representations toward label semantics. Experiments on AAPD and Reuters-21578 demonstrate consistent improvements over strong baselines.</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-branch co-attention with evolving label prototypes for multi-label text classification

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Wentao Lu,
  • Huaxiong Zhang,
  • Lili He,
  • Rougang Zhou

摘要

Multi-label text classification underpins applications such as information retrieval, news tagging, and scholarly annotation, yet existing methods remain vulnerable to noise and redundancy, tend to model label-–text and text-–text relations from a single view, and insufficiently exploit document hierarchies. We propose LTC-MPE, a dual-branch co-attention model with evolving label prototypes. A pretrained encoder first produces contextual token representations, upon which a label-guided word filter constructs a binary mask that sparsifies sequences, suppresses noise. On this basis, two parallel branches derive a label-specific view via label attention and a sentence-level hierarchical view capturing local and global dependencies. A bidirectional co-attention module then deeply aligns and adaptively fuses these complementary representations. Finally, label prototypes are updated online via a momentum mechanism and guide document representations toward label semantics. Experiments on AAPD and Reuters-21578 demonstrate consistent improvements over strong baselines.