【作者】 黄辉；

【导师】 徐金安；

【作者基本信息】 北京交通大学 ， 计算机科学与技术， 2022， 硕士

【摘要】 机器翻译质量评估(Quality Estimation,QE)是研究如何解决在没有参考译文的情况下对机器译文的质量进行评估的问题。机器译文质量评估可以减轻译后编辑工作量,在翻译系统训练时作为译文评价指标,具有重要的学术和实用价值。机器翻译质量评估经历了基于统计的、基于深度学习等方法的发展,近年来,在多语言预训练模型出现之后,由于其强大的特征提取能力,基于预训练模型的翻译质量评估方法成为了主流。但是不断涌现的预训练模型参数越来越多,简单地模型集成会导致推理开销过大,如何在计算资源不足的情况下进行特征融合,成为了一个亟待解决的问题。另外,由于机器翻译质量评估数据标注困难,大部分的训练集仅包含几千到几万条,这导致了严重的数据稀缺问题。针对以上挑战,本文对基于预训练模型的翻译质量评估算法,对低资源场景下特征融合和数据稀缺两方面的问题进行了研究。本研究的贡献主要有以下三点:(1)针对多个预训练模型集成开销过大的问题,本研究提出了迭代集成蒸馏算法,将多个预训练模型所蕴含的知识集成到单个模型中,在不增加推理开销的前提下,大幅提升了翻译质量评估的性能;(2)针对翻译质量评估的数据稀缺问题,本研究提出可以使用对比学习,利用无标注的平行语料来训练翻译质量评估模型,并且提出了使用多源的去噪自编码器来为对比学习构造负例;(3)本研究提出了基于预训练模型的数据增强框架,借助理解式和生成式预训练模型的不同能力,构造了同分布的训练数据,并且借助知识蒸馏进行了解码空间的迁移。本研究在WMT(世界机器翻译大会)、CCMT(全国机器翻译大会)等多个公开数据集上进行了实验验证,证明了所提出的方法的有效性。本研究挖掘了预训练模型在机器翻译质量评估任务上的潜力,提升了在低资源场景下机器翻译质量评估的精度和实用性。更多还原

【Abstract】 Machine Translation Quality Estimation(QE)aims to evaluate the quality of machine translations without reference.QE can reduce the workload of post-translation editing,and can be used as a metric for translation system during training.Previous methods deal with QE by the methods of statistics and deep learning.In recent years,after the emergence of multilingual pre-trained models,pre-trained model-based QE has become mainstream owing to their powerful feature extraction capabilities.However,the pretrained parameters are rapidly increasing,therefore model ensemble will lead to excessive inference overhead.How to integrate different features of multiple models with limited computation resource has become an urgent problem.In addition,due to the difficulty of artificial annotation,most of the QE data only contain thousands of training samples,leading to a serious data sparsity problem.In this paper,we aim to study QE methods in low-resource setting on two problems:feature integration and data sparsity.Our contributions mainly consist of three parts:(1)To cope with the excessive ensemble cost of multiple pre-trained models,we propose an iterative ensemble distillation algorithm,which integrates the knowledge of multiple pre-trained models into a single model.The performance of QE is greatly improved with no additional inference overhead;(2)To deal with the data scarcity problem,we propose to use unlabeled parallel data to train QE models by contrastive learning.We also propose to leverage the multi-source denoising autoencoder to construct negative examples for contrastive learning;(3)We propose a data augmentation framework based on pre-trained models,which leverages different understanding and generative pre-trained models to construct synthetic data with the same distribution,and the decoding space is transferred with the help of knowledge distillation.Experiments on several public datasets such as WMT(Workshop on Machine Translation),CCMT(Chinese Conference on Machine Translation)prove the effectiveness of the proposed methods.Our study taps the potential of the pre-trained model in the task of QE and improves its accuracy and practicability in resource-limited scenarios.更多还原