融合指代消解的神经机器翻译研究  

Neural Machine Translation Combined with Reference Resolution

作　　者：冯勤 贡正仙[1] 李军辉[1] 周国栋[1] FENG Qin;GONG Zhengxian;LI Junhui;ZHOU Guodong(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6期67-76,共10页Journal of Chinese Information Processing

基　　金：国家自然科学基金(61976148)。

摘　　要：篇章中的同一实体经常会呈现出不同的表述,形成一系列复杂的指代关系,这给篇章翻译带来了很大的挑战。该文重点探索指代消解和篇章神经机器翻译的融合方案,首先为指代链设计相应的指代表征;其次使用软约束和硬约束两种方法在翻译系统中实现指代信息的融合。该文建议的方法分别在英语-德语和中文-英语语言对上进行了实验,实验结果表明,相比于同期最好的句子级翻译系统,该方法能使翻译性能获得明显提升。此外,在英语-德语的代词翻译质量的专门评估中,准确率也有显著提升。The same entity in a document often occurs in multiple expressions,forming a series of complex referential relations.To deal with this challenge to Document-level translation,this paper focuses on integrating reference resolution into Document-level machine translation.First,we design the reference representation for the reference chain;Then,we introduce a new translation mechanism to fuse reference representation by combining soft constraint and hard constraint.Experiments on English-German and Chinese-English datasets show that compared with the best sentence-level translation system,our methods can significantly improve translation performance.In addition,experiments show the our methods lead to a strong improvement in the task of evaluating the translation quality of English-German pronouns.

关 键 词：指代表征 神经机器翻译 篇章级机器翻译 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]