![](data:image/png;base64,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)
8.2 基于层次短语的模型 245
真实的情况会更加复杂。对于一个规则的源语言端,可能会有多个不同的目标
语言端与之对应。比如,如下规则的源语言端完全相同,但是译文不同:
X → ⟨ X
1
大幅度 下降 了, X
1
have drastically fallen ⟩
X → ⟨ X
1
大幅度 下降 了, X
1
have fallen drastically ⟩
X → ⟨ X
1
大幅度 下降 了, X
1
has drastically fallen ⟩
输入字符串:
进口 和 出口 大幅度 下降 了
匹配规则:
X → ⟨ X
1
大幅度 下降 了, X
1
have drastically fallen ⟩
X → ⟨ X
1
大幅度 下降 了, X
1
have fallen drastically ⟩
X → ⟨ X
1
大幅度 下降 了, X
1
has drastically fallen ⟩
Span[0,3] 下的翻译假设:
X:imports and exports
S:the import and export
替换 X
1
后生成的翻译假设:
X:imports and exports have drastically fallen
X:the import and export have drastically fallen
X:imports and exports have drastically fallen
X:the import and export have drastically fallen
X:imports and exports has drastically fallen
X:the import and export has drastically fallen
组合
图 8.12 不同规则目标语端及变量译文的组合
这也就是说,当匹配规则的源语言部分“X
1
大幅度 下降 了”时会有三个译文
可以选择。而变量 X
1
部分又有很多不同的局部翻译结果。不同的规则译文和不同的
变量译文都可以组合出一个局部翻译结果。图8.12展示了这种情况的实例。
假设有 n 个规则的源语言端相同,规则中每个变量可以被替换为 m 个结果,对
于只含有一个变量的规则,一共有 nm 种不同的组合。如果规则含有两个变量,这
种组合的数量是 nm
2
。由于翻译中会进行大量的规则匹配,如果每个匹配的源语言
端都考虑所有 nm
2
种译文的组合,解码速度会很慢。
在层次短语系统中,会进一步对搜索空间剪枝。简言之,此时并不需要对所有
nm
2
种组合进行遍历,而是只考虑其中的一部分组合。这种方法也被称作立方剪枝
(Cube Pruning)。所谓“立方”是指组合译文时的三个维度:规则的目标语端、第一
个变量所对应的翻译候选、第二个变量所对应的翻译候选。立方剪枝假设所有的译
文候选都经过排序,比如,按照短语翻译概率排序。这样,每个译文都对应一个坐
标,比如,(i,j,k) 就表示第 i 个规则目标语端、第一个变量的第 j 个翻译候选、第
二个变量的第 k 个翻译候选的组合。于是,可以把每种组合看作是一个三维空间中
的一个点。在立方剪枝中,开始的时候会看到 (0,0, 0) 这个翻译假设,并把这个翻译
假设放入一个优先队列中。之后每次从这个优先队里中弹出最好的结果,之后沿着
三个维度分别将坐标加 1,比如,如果优先队列弹出 (i,j , k),则会生成 (i + 1,j, k)、
(i,j + 1,k) 和 (i,j,k + 1) 这三个新的翻译假设。之后,计算出它们的模型得分,并压