Scholarlink

<p>Neural machine translation (NMT) has achieved remarkable progress by drawing inspiration from the information-processing principles of biological neural systems. In this work, we develop a convolutional-enhanced NMT model that replaces recurrent encoders with multi-layer one-dimensional CNNs, thereby more effectively capturing long-distance dependencies and hierarchical feature abstractions–akin to how the visual cortex hierarchically processes spatial patterns. We systematically explore two convolutional kernel shapes (2 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> 1 and 3 <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> 1) across 3, 6, 10, and 12 layers, and evaluate on the high-quality LDC Chinese–English corpus (1.25 M training sentences) with the NIST02–08 test sets. Our best configuration–a 6-layer network with 3 <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> 1 kernels–yields an average BLEU score of 35.558, representing a +0.99 improvement over the recurrent-baseline. Analysis reveals that 6–10 layers strike the optimal balance between expressive power and over-parameterization, while kernel shape influences the model’s sensitivity to local versus broader context.</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

Biologically inspired convolutional neural architectures for enhanced Chinese–English neural machine translation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月17日

查看全文
  • Zhihao Jiang

摘要

Neural machine translation (NMT) has achieved remarkable progress by drawing inspiration from the information-processing principles of biological neural systems. In this work, we develop a convolutional-enhanced NMT model that replaces recurrent encoders with multi-layer one-dimensional CNNs, thereby more effectively capturing long-distance dependencies and hierarchical feature abstractions–akin to how the visual cortex hierarchically processes spatial patterns. We systematically explore two convolutional kernel shapes (2 \(\times \) 1 and 3 \(\times \) 1) across 3, 6, 10, and 12 layers, and evaluate on the high-quality LDC Chinese–English corpus (1.25 M training sentences) with the NIST02–08 test sets. Our best configuration–a 6-layer network with 3 \(\times \) 1 kernels–yields an average BLEU score of 35.558, representing a +0.99 improvement over the recurrent-baseline. Analysis reveals that 6–10 layers strike the optimal balance between expressive power and over-parameterization, while kernel shape influences the model’s sensitivity to local versus broader context.