Scholarlink

<p>Recently, meta-learning based methods have become prevalent in few-shot text classification, among which Prototypical Network achieves classification by computing the distance between prototypes and query samples in the embedding space. Despite their success, Prototypical Network-based methods often focus on improving the intra-class representation of prototypes while neglecting to capture inter-class distinctions. As a result, when the sampled classes have similar semantics, the prototypes are closely positioned in the embedding space, limiting classification performance. In this paper, we propose a novel Inter-Class Distance Enhanced Prototypical Network (ICDE-PN), which improves generalization by expanding the distances between classes in the embedding space through multi-scale feature learning along the class dimension. In addition, ICDE-PN introduces the Efficient Channel Attention mechanism into few-shot text classification tasks, which selects relevant features while suppressing irrelevant ones to address sample-level noise. Extensive experiments are conducted using four widely recognized benchmark&#xa0;datasets to evaluate our model’s effectiveness. The results reveal that ICDE-PN establishes superiority in 5-shot classification and achieves competitive 1-shot performance. Specifically, on the 20 Newsgroups dataset, ICDE-PN achieves a 3.5% higher accuracy in 5-shot classification compared to the leading baseline method.</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

Inter-class distance enhanced prototypical network for few-shot text classification

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • Yuqi Wang,
  • Cunhe Li

摘要

Recently, meta-learning based methods have become prevalent in few-shot text classification, among which Prototypical Network achieves classification by computing the distance between prototypes and query samples in the embedding space. Despite their success, Prototypical Network-based methods often focus on improving the intra-class representation of prototypes while neglecting to capture inter-class distinctions. As a result, when the sampled classes have similar semantics, the prototypes are closely positioned in the embedding space, limiting classification performance. In this paper, we propose a novel Inter-Class Distance Enhanced Prototypical Network (ICDE-PN), which improves generalization by expanding the distances between classes in the embedding space through multi-scale feature learning along the class dimension. In addition, ICDE-PN introduces the Efficient Channel Attention mechanism into few-shot text classification tasks, which selects relevant features while suppressing irrelevant ones to address sample-level noise. Extensive experiments are conducted using four widely recognized benchmark datasets to evaluate our model’s effectiveness. The results reveal that ICDE-PN establishes superiority in 5-shot classification and achieves competitive 1-shot performance. Specifically, on the 20 Newsgroups dataset, ICDE-PN achieves a 3.5% higher accuracy in 5-shot classification compared to the leading baseline method.