摘要
针对目前主流的神经机器翻译模型Transformer内部结构单元堆叠而造成的底层信息丢失和多层单元输出信息偏差不同的问题,对其结构进行了改进,提出了一种融合底层信息的神经机器翻译模型。采用多种网络结构对源语言进行底层信息的特征提取,并采用残差连接的方式实现底层信息的向上传递。实验结果显示:融合底层信息后的翻译模型在电气工程领域内的双语评估研究(BLEU)值最多提升了2.47个百分点。
Since the development of neural machine translation,it has achieved good translation performance in the general field,but has not achieved good results in the field of low resources.In this paper,the current mainstream neural machine translation model,Transformer,has improved its structure to solve the problem of the loss of the underlying information caused by the stacking of cells in its internal structure and the different deviation of the output information of multi-level cells.A neural machine translation model that integrates the underlying information was proposed.The multiple network structures were used to extract the features of the underlying information of the source language,and the way of residual connection was used to realize the upward transmission of the underlying information.The experimental results show that the translation model fused with underlying information has a maximum increase of 2.47 percentage points in bilingual evaluation understudy(BLEU)values in the field of electrical engineering.
作者
陈媛
陈红
CHEN Yuan;CHEN Hong(School of Foreign Languages,Henan University of Science and Technology,Luoyang 471023,China;College of Information Engineering,Henan University of Science and Technology,Luoyang 471023,China;Henan Province Electronics and Power Devices and Systems Research Center,Henan University of Science and Technology,Luoyang 471023,China)
出处
《河南科技大学学报（自然科学版）》
CAS
北大核心
2023年第6期42-48,M0004,M0005,共9页
Journal of Henan University of Science And Technology:Natural Science
基金
国家自然科学基金项目(U2004163)。
