融合多特征的汉维神经网络机器翻译模型

针对汉维神经网络机器翻译中出现的未登录词过多、维吾尔语端形态生成以及汉维词语表意不一致等问题,提出一种融合"编码器-解码器"特征、维吾尔语"词干-词缀"语言模型特征、汉维-维汉双向词对齐特征的汉维翻译策略。综合考虑汉维语言差异、汉维语言资源稀缺等问题,将统计机器翻译中的双语知识引入到神经网络机器翻译模型中,多个特征通过一个对数线性模型组合。实验结果表明,该方法能够有效提升汉维神经网络机器翻译性能,平均BLEU提升大于2.0。

Optimized Chinese-Uyghur neural machine translation model based on multi-features

To overcome OOVs,morphological generation and n-to-1 alignment exist in Chinese-Uyghur neural machine translation,a log-linear model based Chinese-Uyghur machine translation approach was proposed,which integrated features like encoder-decoder feature,Uyghur stem-affix language model feature and Chinese-Uyghur bidirectional word alignment feature.Knowledge of statistical machine translation was imported into neural machine translation,which alleviated data sparsity and morphological differences during model training.Experimental results show that the proposed model can achieve 2.0+ BLEU improvements compared with other methods.