融合无监督质量估计指标的译文质量估计方法

Quality Estimation Method Incorporating Unsupervised Quality Estimation Indicators

质量估计的目的是在没有参考译文的情况下衡量翻译内容的质量,这对于需要高质量翻译任务中的机器翻译系统至关重要.针对有监督的翻译质量估计中普遍存在的缺乏标记训练数据和模型框架中两阶段学习目标差异的问题,以及无监督的质量估计中存在的因为任务目标模糊、特征学习不完全而导致最终结果远不如有监督模型的问题,本文提出了将无监督质量估计中的估计指标当作特征融入有监督的质量估计模型中的方法,以此来互相弥补两种模型之间存在的缺点.在WMT2020的高资源对和低资源对上的实验结果证实,相对于基线的有监督和无监督模型,两者结合的方法能够更好的提高翻译质量估计的准确性,与人工评分的皮尔逊相关系数都有所提升.

The purpose of quality estimation is to measure the quality of translated content without reference translation,which is very important for machine translation systems that need high-quality translation tasks.In view of the widespread lack of labeled training data and the difference between the two-stage learning objectives in the model framework in supervised translation quality estimation,as well as the problem that the final result of unsupervised translation quality estimation is far inferior to that of supervised model because of vague task objectives and incomplete feature learning,this paper puts forward a method of integrating the estimation indicators in unsupervised translation quality estimation into supervised quality estimation model as features,so as to make up for the shortcomings between the two models.The experimental results on WMT2020′s high-resource pair and low-resource pair show that,compared with the baseline supervised and unsupervised models,the combination of the two methods can better improve the accuracy of translation quality estimation,and the Pearson correlation coefficient with manual evaluation has been improved.