![](data:image/png;base64,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)
270 Chapter 8. 基于句法的模型 肖桐 朱靖波
引入 d
′
的意义在于,整个分析过程具有了递归性。从超图上看,d
′
可以对应以
一个(或几个)节点为“根”的子图,因此只需要在这个(或这些)子图上增加新的
超边就可以得到更大的推导。这个过程不断执行,最终完成对整个句子的分析。
在句法分析中,超图的结构往往被组织为一种表格(Chart)结构。表格的每个
单元代表了一个跨度,因此可以把所有覆盖这个跨度的推导都放入相应的表格单元
(Chart Cell)。对于上下文无关文法,表格里的每一项还会增加一个句法标记,用来
区分不同句法功能的推导。
cell[1,2] cell[0,2]
cell[0,1]
N/A
VV[0,1]
NN[1,2]
NP[1,2]
VP[0,2]
NP[0,2]
跨度大小
起
始
位
置
Chart(表格)
图 8.38 句法分析 Chart(表格)结构实例
如图8.38所示,覆盖相同跨度的节点会被放入同一个表格单元,但是不同句法标
记的节点会被看作是不同的项(Item)。这种组织方式建立了一个索引,通过索引可
以很容易地访问同一个跨度下的所有推导。比如,如果采用自下而上的分析,可以
从小跨度的表格单元开始,构建推导,并填写表格单元。这个过程中,可以访问之前
的表格单元来获得所需的局部推导(类似于前面提到的 d
′
)。该过程重复执行,直到
处理完最大跨度的表格单元。而最后一个表格单元就保存了完整推导的根节点。通
过回溯的方式,能够把所有推导都生成出来。
基于句法的机器翻译仍然可以使用超图进行翻译推导的表示。和句法分析一样,
超图的每条边可以对应一个基于树结构的文法,超边的头代表文法的左部,超边的
尾代表规则中变量所对应的超图中的节点
10
。图8.39 给出了一个使用超图来表示机
器翻译推导的实例。可以看到,超图的结构是按源语言组织的,但是每个规则(超
边)会包含目标语言的信息。由于同步翻译文法可以确保规则的源语言端和目标语
言端都覆盖连续的词串,因此超图中的每个节点都对应一个源语言跨度,同时对应
一个目标语的连续译文。这样,每个节点实际上代表了一个局部的翻译结果。
不过,机器翻译与句法分析也有不同之处。最主要的区别在于机器翻译使用了
语言模型作为一个特征,比如 n-gram 语言模型。因为语言模型并不是上下文无关的,
因此机器翻译中计算最优推导的方法和句法分析会有不同。常用的方法是,直接在
每个表格单元中融合语言模型的分数,保留前 k 个结果;或者,在构建超图时不计算
语言模型得分,等到构建完整个超图之后对最好的若干个推导用语言模型重新排序;
10
也可以把每个终结符看作是一个节点,这样一个超边的尾就对应规则的树片段中所有的叶子。