利用词链提高文档级神经机器翻译的词汇翻译一致性  

Utilizing Word Link to Improve Lexical Translation Consistency for Document-Level Neural Machine Translation

作　　者：雷翔宇 李军辉[1] LEI Xiangyu;LI Junhui(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》2025年第2期72-79,共8页Journal of Chinese Information Processing

基　　金：江苏高校优势学科建设工程资助项目。

摘　　要：近年来,各种上下文感知模块的引入,使得文档级神经机器翻译(Document-level Neural Machine Translation,DNMT)取得了令人瞩目的进步。受“一个语篇一个翻译(one translation per discourse)”的启发,该文在代表性DNMT模型G-Transformer的基础上,提出一种有效的方法对源端文档中重复出现的词汇进行建模以缓解词汇翻译不一致问题。具体来说,首先获取源端文档中每个单词的词链;然后,使用词链注意力机制以交换同一词链单词之间的上下文信息,从而增强词汇翻译一致性。基于汉?英和德?英文档级翻译任务的实验结果表明,该文的办法不仅显著缓解了词汇翻译不一致的问题,而且提高了翻译性能。In recent years,various context-aware modules has made remarkable progress in Document-level Neural Machine Translation.Inspired by"one translation per discourse",this paper propose an effective method to model repeated words in source-side documents to alleviate the problem of the lexical translation inconsistency.Based on G-Transformer,a representative DNMT model,we first obtain word links for each word in source-side documents.Then,we propose the word-link-attention to exchange context information between words in the same word link to enhance the consistency of lexical translation.Experimental results on Chinese English and German English document-level translation tasks show the effectiveness of our approach.

关 键 词：文档级神经机器翻译 词汇翻译一致性 词链 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]