基于多语言预训练语言模型的译文质量估计方法

(1.中国科学院自动化研究所,模式识别国家重点实验室,北京 100190; 2.中国科学院大学,北京 100049)

机器翻译; 译文质量估计; 深度学习

Quality estimation based on multilingual pre-trained language model
LU Jinliang1,2,ZHANG Jiajun1,2*

(1.National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China; 2.University of Chinese Academy of Sciences,Beijing 100049,China)

DOI: 10.6043/j.issn.0438-0479.201908034

备注

传统的机器翻译评价方法往往需要参考译文,利用机器双语互译评估(BLEU)值等方法比较翻译结果与参考译文之间的相似性.但是,在现实生活中却很难为每一句待翻译的句子找到参考答案,因此,不使用参考译文的译文质量估计(quality estimation,QE)方法有着更加广泛的应用场景.在该文中,基于多语言的预训练语言模型,利用联合编码的策略完成句子级的QE任务,在WMT 2018的QE任务德语→英语语言方向上的评测数据集上取得了最佳的实验结果.同时,对比了微调过程中不同网络结构对于该任务的影响,并探究了平行语料联合编码二次预训练在句子级跨语言任务上的效果.

In recent years,neural machine translation has advanced greatly.Traditional machine translation evaluation methods generally require references,such as BLEU(bilingual evaluation understudy).These methods aim to compare similarities between candidate and reference.However,in practice,it is difficult for us to find a reference for each source sentence.Therefore,the quality estimation(QE)application scenario is more extensive.In this paper,we use the multi-language pre-trained language model,with the joint-encoding strategy to complete the sentence-level QE task.Experiments show that our model can obtain outstanding results in WMT 2018 QE Shared Task German→English language direction.At the same time,we also compare the impact of different network structures on the task.Finally,we explore the effect of the secondary pre-trained of parallel corpus on the cross-lingual sentence tasks.