一种面向统计机器翻译的协同权重训练方法

摘要

分析了统计机器翻译中的特征权重的领域自适应问题,并针对该问题提出了协同的权重训练方法.该方法使用来自不同解码器的译文作为准参考译文,并将其加入到开发集中,使得特征权重的训练过程向测试集所在的领域倾斜.此外,提出了使用最小贝叶斯风险的系统融合方法来选择准参考译文,进一步提高了协同权重训练的性能.实验结果表明,使用最小贝叶斯风险系统融合的协同训练方法,可以在一定程度上解决特征权重的领域自适应问题,并显著地提高了在目标领域内机器翻译结果的质量.%In this paper, based on the investigation of domain adaptation for feature weight, the study proposes to use a co-training framework to handle domain adaptation for feature weight, i.e. The study uses the translation results from another heterogeneous decoder as pseudo references and adds them to the development data set for minimum error rate training to bias the feature weight to the domain of test data set. Furthermore, the study uses a minimum Bayes- Risk combination for pseudo reference selection, which can pick proper translation results from the translation candidates from both decoders to smooth the training process. Experimental results show that this co-training method with a minimum Bayes-Risk combination can yield significant improvements in target domain.