面向汉维机器翻译的双语关联度优化模型

Bilingual relatedness optimization model for Chinese-Uyghur machine translation

针对汉语—维吾尔语的统计机器翻译系统中存在的语义无关性问题,提出基于神经网络机器翻译方法的双语关联度优化模型。该模型利用注意力机制捕获词对齐信息,引入双语短语间的语义相关性和内部词汇匹配度,预测双语短语的生成概率并将其作为双语关联度,以优化统计翻译模型中的短语翻译得分。在第十一届全国机器翻译研讨会(CWMT 2015)汉维公开机器翻译数据集上的实验结果表明,与基线系统相比,在使用较小规模的训练数据和词汇表的条件下,所提方法可以同时有效地提高短语级别和句子级别的机器翻译任务性能,分别获得最高2.49和0.59的BLEU值提升。

Focused on the issue of semantic independence in Chinese-Uyghur statistical machine translation system,this paper proposed a bilingual relatedness optimization model based on neural machine translation method.The model utilized the attention mechanism to capture word alignment information as well as introduced bilingual phrase semantic relevance and inner word correlation to predict the conditional probability of bilingual phrase pair.And it took the probability as bilingual relatedness to optimize the phrase translation scores in statistical translation model.Experimental results on the 11th China Workshop on Machine Translation(CWMT 2015)Chinese-Uyghur public machine translation datasets show that the proposed approach can achieve obvious improvements both in the phrase-level and the sentence-level machine translation tasks,which outperforms the baseline system with a relative small-scale training data and vocabulary.The highest BLEU point of the proposed algorithm gains 2.49 and 0.59 respectively.