![](data:image/png;base64,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)
230 Chapter 8. 基于句法的模型 肖桐 朱靖波
进口
在 过去 的 五 到 十 年 间
有了 大幅度 下降
The imports
drastically fell in the past five to ten years
图 8.1 汉英翻译中的不同距离下的依赖
际系统研发中仍然不现实。图8.1展示了一个汉语到英语的翻译实例。源语言的两个
短语(蓝色和红色高亮)在目标语言中产生了调序。但是,这两个短语在源语言句子
中横跨 8 个单词。如果直接使用这 8 个单词构成的短语进行翻译,显然会有非常严
重的数据稀疏问题,因为很难期望在训练数据中见到一模一样的短语。
仅使用连续词串不能处理所有的翻译问题,其根本原因在于句子的表层串很难
描述片段之间大范围的依赖。一个新的思路是使用句子的层次结构信息进行建模。第
三章已经介绍了句法分析基础。对于每个句子,都可以用句法树描述它的结构。
S
VP
ADVP
VBN
fallen
RB
drastically
VBZ
have
NP
NP
this area
IN
in
NP
NN
imports
DT
The
图 8.2 一棵英语句法树(短语结构树)
图8.2就展示了一棵英语句法树(短语结构树)。句法树描述了一种递归的结构,
每个句法结构都可以用一个子树来描述,子树之间的组合可以构成更大的子树,最
终完成整个句子的表示。相比线性的序列结构,树结构更容易处理大片段之间的关
系。比如,两个在序列中距离“很远”的单词,在树结构中可能会“很近”。
句法树结构可以赋予机器翻译对语言进一步抽象的能力,这样,可以不需要使
用连续词串,而是通过句法结构来对大范围的译文生成和调序进行建模。图8.3是一
个在翻译中融入源语言(汉语)句法信息的实例。这个例子中,介词短语“在 ... 后”
包含 11 个单词,因此,使用短语很难涵盖这样的片段。这时,系统会把“在 ... 后”
错误地翻译为“In ...”。通过句法树,可以知道“在 ... 后”对应着一个完整的子树结
构 PP(介词短语)。因此也很容易知道介词短语中“在 ... 后”是一个模板(红色),
而“在”和“后”之间的部分构成从句部分(蓝色)。最终得到正确的译文“After ...”。
使用句法信息在机器翻译中并不新鲜。在基于规则和模板的翻译模型中,就大