摘要
神经机器翻译模型基本都是编码器和解码器结构,在进行编码之前都需要把文字抽象成数学符号,以词向量的形式传入编码器。大多数对词向量内部信息进行处理的方法都需要用到大量的资源进行训练,并不能够简单而快速的提升机器翻译性能。针对此问题,基于Transformer模型,提出了一种利用门控循环单元(gated recurrent unit,GRU)对词向量过滤的技术,直接用训练数据进行操作。首先将词向量和位置向量相加后传入GRU,把GRU对词向量过滤后的输出与词向量进行拼接,然后把拼接后的特征向量经过线性层深度融合在一起,这样得到的词向量就是经过门控单元过滤后的词向量表征,最后再传入编码器和解码器。实验结果表明:在Multi30k和IWSLT2016这2个数据集和不同的模型中,提出的方法都能够使机器翻译的BLEU值得到提升,取得较好的翻译性能。
ed into mathematical notation,which are passed into the encoder in the form of word vectors.Most of the methods for processing the internal information of word vectors need a lot of resources for training,which cannot simply and quickly improve the performance of machine translation.To address this issue,a technique using gated recurrent unit(GRU)to filter word vectors based on the Transformer model is proposed,which directly operates on training data.Firstly,the word vector and position vector are added and transmitted to the GRU.The output filtered by the GRU is concatenated with the word vector.Then,the concatenated feature vectors are deeply fused together through a linear layer.The resulting word vector is represented by the word vector filtered by the gating unit,and finally transmitted to the encoder and decoder.The experimental results show that in the two data sets of Multi30k and IWSLT2016 and different models,the proposed method can make the BLEU of machine translation worth improving and achieve better translation performance.
作者
高帅
周红志
康卫
GAO Shuai;ZHOU Hongzhi;KANGWei(School of Computer and Information Engineering,Fuyang Normal University,Fuyang Anhui 236037,China;College of Information Engineering,Fuyang Normal University,Fuyang Anhui 236041,China)
出处
《阜阳师范大学学报（自然科学版）》
2024年第2期38-43,共6页
Journal of Fuyang Normal University:Natural Science
基金
安徽省高校自然科学重点项目(2022AH052821)。
作者简介
高帅(1997-),男,硕士生,研究方向:机器翻译;通讯作者:周红志(1981-),女,教授,研究方向:大数据分析与数据挖掘,Email:40696670@qq.com.
