Scholarlink

<p>Text classification is a foundational natural language processing task. Many models have transformed text data into innovative graph structures and employed graph neural networks (GNNs) to learn representations for classification. However, graphs constructed based on artificial rules may contain redundant connections, which can introduce noise. Additionally, graph neural network-based (GNN-based) models are ineffective in learning word order. To cope with these difficulties, we propose graph affine Transformer with a symmetric adaptation strategy (GATSAS) for text classification. After graph construction, a symmetric gated graph neural network is crafted to detect edge attributes and remove trivial connections. In the graph learning step, graph affine Transformer encodes word positions and updates word representations with multi-head graph affine attention. Then, the readout module fuses word features to achieve text representation. The experimental results show that GATSAS has an average accuracy of 0.11 <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_6992_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(-\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>-</mo> </math></EquationSource> </InlineEquation>2.37% higher than the GNN-based benchmark models.</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

Graph affine Transformer with a symmetric adaptation strategy for text classification

  • 期刊论文

  • 发表日期：

    2025年2月14日

查看全文
  • Minyi Ma,
  • Hongfang Gong,
  • Yingjing Ding

摘要

Text classification is a foundational natural language processing task. Many models have transformed text data into innovative graph structures and employed graph neural networks (GNNs) to learn representations for classification. However, graphs constructed based on artificial rules may contain redundant connections, which can introduce noise. Additionally, graph neural network-based (GNN-based) models are ineffective in learning word order. To cope with these difficulties, we propose graph affine Transformer with a symmetric adaptation strategy (GATSAS) for text classification. After graph construction, a symmetric gated graph neural network is crafted to detect edge attributes and remove trivial connections. In the graph learning step, graph affine Transformer encodes word positions and updates word representations with multi-head graph affine attention. Then, the readout module fuses word features to achieve text representation. The experimental results show that GATSAS has an average accuracy of 0.11 \(-\) - 2.37% higher than the GNN-based benchmark models.