融合目标端上下文的篇章神经机器翻译  

Modeling Target-side Context for Document-level Neural Machine Translation

作　　者：贾爱鑫 李军辉[1] 贡正仙[1] 张民[1] JIA Aixin;LI Junhui;GONG Zhengxian;ZHANG Min(School of Computer Science and Technology,Soochow University,Suzhou,Jiangsu 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《中文信息学报》2024年第4期59-68,共10页Journal of Chinese Information Processing

基　　金：国家自然科学基金(61876120,61976148)。

摘　　要：神经机器翻译在句子级翻译任务上取得了令人瞩目的效果,但是句子级翻译的译文会存在一致性、指代等篇章问题,篇章翻译通过利用上下文信息来解决上述问题。不同于以往使用源端上下文建模的方法,该文提出了融合目标端上下文信息的篇章神经机器翻译。具体地,该文借助推敲网络的思想,对篇章源端进行二次翻译,第一次基于句子级翻译,第二次翻译参考了全篇的第一次翻译结果。基于LDC中英篇章数据集和WMT英德篇章数据集的实验结果表明,在引入较少的参数的条件下,该文方法能显著提高翻译性能。同时,随着第一次翻译(即句子级译文)质量的提升,所提方法也更有效。Recently neural machine translation(NMT)has achieved remarkable success in sentence-level translation.However,it still cannot resolve a wide variety of discourse phenomena,such as lexical cohesion and coreference,which can be alleviated by using context information in document-level translation.In contrast to existing studies of modeling source-side context,this paper proposes to model target-side context in document-level NMT.Specifically,motivated by deliberation networks,our approach translates source-side document twice.In the first-pass translation,it performs sentence-level translation.In the second-pass,it properly translates each sentence by modeling the target-side context which has just be generated from the first-pass translation.Experimental results on LDC Chinese-to-English and WMT English-to-German document-level translation tasks show that our approach significantly improves translation performance by introducing few parameters.Meanwhile,it is observed that the proposed approach benefits more if the performance of the first-pass translation(i.e.,sentence-level NMT)is improved.

关 键 词：神经机器翻译 推敲网络 篇章翻译 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]