《厦门大学学报（自然科学版）》

目前性能最优的译文质量估计系统使用神经机器翻译中的编码器-解码器模型作为特征提取器.该方法由于限制词表大小易导致数据稀疏问题,从而使得较多的未登陆词不能被正确评价.为了缓解上述问题,在详细分析不同子词切分方法的特点后,提出了基于字节对编码(BPE)子词切分和基于一元文法语言模型子词切分的神经译文质量估计方法,并将两者的译文质量估计的得分与基于词语切分的神经译文质量估计得分融合后进行译文质量估计.在WMT18句子级别译文质量估计子任务数据集上的实验结果表明:融合BPE子词切分、一元文法语言模型子词切分和词语切分的神经译文质量估计方法的性能在多个评测子任务上超过了WMT18给出的最好参与系统,深入的实验分析进一步揭示了融合不同粒度的句子切分方法提高了译文质量估计的健壮性.

Nowadays, the state-of-the-art translation quality estimation system takes the encoder-decoder model in neural machine translation as feature extractor.Duing to the restriction of vocabulary size,this method is prone to data sparseness,so that many out-of-vocabulary words can't be correctly evaluated.To tackle the data sparse issues,we propose the neural quality estimation approaches based on the byte-pair-encoding(BPE)subword unit and unigram language model subword unit after a detailed discussion of the characteristic of different subword segmenters.Furthermore,results of the neural quality estimation systems based on BPE subwords and that of the unigram language model subwords are combined with the results of the neural quality estimation systems based on words.Experimental results on the data sets of WMT18 sentence-level quality estimation tasks show that the ensemble system combining the results of the neural quality estimation systems based on BPE subwords,unigram language model subwords and words perform better than the best participated systems on several translation directions in WMT18 quality estimation task.Deep analyses further reveal that the ensemble system combining the results of neural quality estimation systems based on different granularity segmentations improve the robustness of the quality estimation system.

引言
1 UNQE模型
2 子词切分
3 实验
4 结论

图1 UNQE模型结构图[16]<br/>Fig.1 An illustration of the UNQE model architecture[16]

图1 UNQE模型结构图[16]
Fig.1 An illustration of the UNQE model architecture[16]

表1 使用不同切分方法的结果示例<br/>Tab.1 An example of the sentence segmentation results by different methods

表1 使用不同切分方法的结果示例
Tab.1 An example of the sentence segmentation results by different methods

表2 编码器-解码器子网络预训练使用的双语平行语料及规模<br/>Tab.2 The bilingual corpus and its size used to pre-train the encoder-decoder sub-network

表2 编码器-解码器子网络预训练使用的双语平行语料及规模
Tab.2 The bilingual corpus and its size used to pre-train the encoder-decoder sub-network

表3 使用不同切分方法的统计词表大小<br/>Tab.3 The statistics of the vocabulary size when sentences segmented by different segmentation method

表3 使用不同切分方法的统计词表大小
Tab.3 The statistics of the vocabulary size when sentences segmented by different segmentation method

表4 不同模型在英语-捷克语和英语-拉脱维亚语翻译任务中的性能<br/>Tab.4 The quality estimation performances of different models on the English-Czech and English-Latvian data set transtation tasks

表4 不同模型在英语-捷克语和英语-拉脱维亚语翻译任务中的性能
Tab.4 The quality estimation performances of different models on the English-Czech and English-Latvian data set transtation tasks

表5 不同模型在英语-德语和德语-英语翻译任务中的性能<br/>Tab.5 The quality estimation performances of different models on the English-German and German-English data set translation tasks

表5 不同模型在英语-德语和德语-英语翻译任务中的性能
Tab.5 The quality estimation performances of different models on the English-German and German-English data set translation tasks

表6 不同子词切分方法对译文质量估计影响示例<br/>Tab.6 Examples of the effects of different segmentation methods on quality estimation

表6 不同子词切分方法对译文质量估计影响示例
Tab.6 Examples of the effects of different segmentation methods on quality estimation

[1] SPECIA L,SHAH K,DE SOUZA J G,et al.QuEst-a translation quality estimation framework[C]∥Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics:System Demonstrations.Sofia:ACL,2013:79-84.
[2] MIKOLOV T,KARAFIÁT M,BURGET L,et al.Recurrent neural network based language model[C]∥Eleventh Annual Conference of the International Speech Communication Association.Makuhari:DBLP,2010:1045-1048.
[3] MIKOLOV T,SUTSKEVER I,CHEN K,et al.Distri-buted representations of words and phrases and their compositionality[C]∥Advances in Neural Information Processing Systems.Lake Tahoe:NIPS,2013:3111-3119.
[4] BAHDANAU D,CHO K,BENGIO Y.Neural machine translation by jointly learning to align and translate[EB/OL].[2019-08-01].http:∥arxiv.org/pdf/1409.0473.pdf.
[5] SENNRICH R,FIRAT O,CHO K,et al.Nematus:a toolkit for neural machine translation[EB/OL].[2019-08-01].http:∥arxiv.org/pdf/1703.04357.pdf.
[6] VASWANI A,SHAZEER N,PARMAR N,et al.Attention is all you need[EB/OL].[2019-08-01].http:∥arxiv.org/pdf/1606.02003.pdf.
[7] 李亚超,熊德意,张民.神经机器翻译综述[J].计算机学报,2018,41(12):100-121.
[8] 刘洋.神经机器翻译前沿进展[J].计算机研究与发展,2017:54(6):1144-1149.
[9] SHAH K,NG RAVMOND W M,BOUGARES F,et al.Investigating continuous space language models for machine translation quality estimation[C]∥Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing.Lisbon:EMNLP,2015:1073-1078.
[10] CHEN Z,TAN Y,ZHANG C,et al.Improving machine translation quality estimation with neural network features[C]∥Proceedings of the Second Conference on Machine Translation.Copenhagen:EMNLP,2017:551-555.
[11] 陈志明,李茂西,王明文.基于神经网络特征的句子级别译文质量估计[J].计算机研究与发展,2017,54(8):1804-1812.
[12] SPECIA L,BLAIN F,LOGACHEVA V,et al.Findings of the WMT 2018 shared task on quality estimation[C]∥Proceedings of the Third Conference on Machine Translation:Shared Task Papers.Belgium:ACL,2018:689-709.
[13] KIM H,JUNG H Y,KWON H,et al.Predictor-estimator:neural quality estimation based on target word prediction for machine translation[J].ACM Transactions on Asian and Low-Resource Language Information Processing(TALLIP),2017,17(1):3.
[14] KIM H,LEE J H.A recurrent neural networks approach for estimating the quality of machine translation output[C]∥Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics:Human Language Technologies.San Diego:NAACL,2016:494-498.
[15] KIM H,LEE J H,NA S H.Predictor-estimator using multilevel task learning with stack propagation for neural quality estimation[C]∥Proceedings of the Second Conference on Machine Translation.Copenhagen:EMNLP,2017:562-568.
[16] LI M X,XIANG Q Y,CHEN Z M,et al.A unified neural network for quality estimation of machine translation[J].IEICE Transactions on Information and Systems,2018,101(9):2417-2421.
[17] FAN K,WANG J Y,Li B,et al."Bilingual expert" can find translation errors[C]∥Proceedings of the AAAI Conference on Artificial Intelligence.Hawaii:AAAI,2019:6367-6374.
[18] SENNRICH R,HADDOW B,BIRCH A.Neural machine translation of rare words with subword units[EB/OL].[2019-08-01].http:∥arxiv.org/pdf/1508.07909.pdf.
[19] KUDO T.Subword regularization:improving neural network translation models with multiple subword candidates[EB/OL].[2019-08-01].http:∥arxiv.org/pdf/1804.10959.pdf.

备注

引言

1 UNQE模型

2 子词切分

3 实验

4 结论

学报简介

备注

引言

1 UNQE模型

2 子词切分

3 实 验

4 结 论

学报简介

3 实验

4 结论