![](data:image/png;base64,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)
180 Chapter 6. 基于扭曲度和繁衍率的模型 肖桐 朱靖波
6.1.1 什么是扭曲度
调序(Reordering)是自然语言翻译中特有的语言现象。造成这个现象的主要原
因在于不同语言之间语序的差异,比如,汉语是“主谓宾”结构,而日语是“主宾
谓”结构。即使在句子整体结构相似的语言上进行翻译,调序也是频繁出现的现象。
如图6.1所示,当一个主动语态的汉语句子翻译为一个被动语态的英语句子时,如果
直接顺序翻译,那么翻译结果“I with you am satisfied”很明显不符合英语语法。这
里就需要采取一些方法和手段在翻译过程中对词或短语进行调序,从而得到正确的
翻译结果。
源语
我
对
你
感到
满意
顺序翻译
I
with
you
am
satisfied
(a) 顺序翻译对齐结果
源语
我
对
你
感到
满意
调序翻译
I
am
satisfied with
you
(b) 调序翻译对齐结果
图 6.1 顺序翻译和调序翻译的实例对比
在对调序问题进行建模的方法中,最基本的是使用调序距离方法。这里,可以
假设完全进行顺序翻译时,调序的“代价”是最低的。当调序出现时,可以用调序相
对于顺序翻译产生的位置偏移来度量调序的程度,也被称为调序距离。图6.2展示了
翻译时两种语言中词的对齐矩阵。比如,在图6.2(a) 中,系统需要跳过“对”和“你”
来翻译“感到”和“满意”,之后再回过头翻译“对”和“你”,这样就完成了对单词
的调序。这时可以简单地把需要跳过的单词数看作一种距离。
可以看到,调序距离实际上是在度量目标语言词序相对于源语言词序的一种扭
曲程度。因此,也常常把这种调序距离称作扭曲度(Distortion)。调序距离越大对应
的扭曲度也越大。比如,可以明显看出图6.2(b) 中调序的扭曲度要比图6.2(a) 中调序
的扭曲度大,因此6.2(b) 实例的调序代价也更大。
在机器翻译中使用扭曲度进行翻译建模是一种十分自然的想法。接下来,会介
绍两个基于扭曲度的翻译模型,分别是 IBM 模型 2 和隐马尔可夫模型。不同于 IBM