多机制合并注意力的多路径神经机器翻译方法
DOI:
CSTR:
作者:
作者单位:

(昆明理工大学 信息工程与自动化学院,云南 昆明 650500)

作者简介:

范洪博(1982-),男,硕士生导师,主要研究方向为机器翻译与区块链.

通讯作者:

中图分类号:

基金项目:


A multi-path neural machine translation method based on multi-mechanism atten tion
Author:
Affiliation:

(Faculty of Information Engineering and Automation,Kunming University of Scienc e and Technology,Kunming 650500,China)

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    注意力机制是目前神经机器翻译的主流技术,目 前已提出了多种注意力生成机制,各 机制生成的注意力各有优劣,但每种机制都不能充分利用全部已知信息,其结果和理论上的 真实注意力具有一定差距,影响翻译质量。本文提出一种基于民主决策的合并注意力生成方 法,将多种注意力生成机制所产生的注意力进行加权叠加与归一化后,所生成的值作为新的 注意力,用于指导解码器的翻译过程。类似民主决策会比独断专行的决策获得更好的决策准 确度,该机制可以获得相对更为准确的注意力,进而提升翻译质量。基于上述方法,本文在 CNN、Transformer、Tree Transformer三个算法的基础上,生成合并注意力,提出MA-CTT 算 法,在开放德英语料(IWSLT14)上,MA-CTT获得了32.61的BLEU, 翻译准确度明显高于各基础算法。

    Abstract:

    Attention based neural network translation is the mainstream technology in the field of machine translation.At the moment,a variety of attention gene rating mechanisms has been proposed.Although these attention generation mechani sms have their own advantages and disadvantages,each kind of attention generati on mechanism cannot make full use of all the known information.Every attention generation mechanism has a certain gap with the theoretical real attention.In t his paper,a variety of attention generation mechanisms are used to generate the ir own attention,and then these attentions are weighted and normalized to guide the decoder for further translation process.Similar to the democratic decision -making process which achieves better decision-making accuracy than the arbitr ar y ones,this mechanism can obtain relatively more accurate attention value,thus improving the overall quality of translation.Based on the above algorithms,th is paper designs a new algorithm MA-CTT using CNN,Transformer and Tree Transfor m er to generate a combined attention on the open German English corpus(IWSLT14). This paper combines the advantages of CNN,Transformer and Tree Transformer.Acc ording to the experimental data,the BLEU value of the algorithm reaches 32.61o n the open German English corpus (IWSLT14).Compared with the baseline algorithms ,the translation accuracy of this algorithm has been improved.

    参考文献
    相似文献
    引证文献
引用本文

范洪博,郑棋.多机制合并注意力的多路径神经机器翻译方法[J].光电子激光,2021,32(5):491~498

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2020-12-02
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2021-05-28
  • 出版日期:
文章二维码