机器翻译系统特征权值的贝叶斯优化方法

Feature weights Bayesian optimization method of machine translation system

针对机器翻译中存在的特征权重的领域自适应问题,提出一种联合最小贝叶斯融合的系统权重训练方法。在协同训练的框架内,采用不同解码器的输出作为参考译文,通过扩展开发集,保证特征权重训练的有效性。通过利用最小贝叶斯风险融合方法,提升协同训练的稳定性。实验结果表明,该方法较好解决了特征权重的领域自使用问题,优化了机器翻译质量。

To solve the domain adaptive problem of feature weight in machine translation,a method of weight training based on joint minimum bias fusion was proposed.In the framework of cooperative training,the output of different decoders was used as the reference version,and the effectiveness of feature weight training was guaranteed by extending the development set.The method improved the stability of cooperative training by using the minimum Bias risk fusion method.Experimental results show that the proposed method can better solve the problem of domain self usage of feature weight and optimize the quality of machine translation.