汉维统计机器翻译中的形态学处理

Morphology Processing in Chinese-Uyghur Statistical Machine Translation

针对汉语和维吾尔语语序差别(前者是主-谓-宾结构,后者是主-宾-谓结构)及形态差别较大的问题,通过编写调序规则将汉语调整为主-宾-谓结构,将维吾尔语单词切分为词干、词缀等更小的词素单元来训练统计模型,同时测试词素的切分粒度对翻译性能的影响。实验结果表明,对汉语句法结构的调整及以词干、词缀等更小的词素形式参与训练可以有效提高翻译质量。

For the large differences of syntactic structure between Chinese and Uyghur, it composes rules to reorder the structure of Chinese sentences to that of Uyghar. For the large morphological differences between Chinese and Uyghur, it splits Uyghur words into stems and affixes, that is, morphemes, to train the statistical model. Meanwhile, it tests the effects of splitting granularities on translation performance. Experimental results show Chinese sentence reordering and splitting Uyghur words into morphemes can effectively improve the performance of translation system.