![](data:image/png;base64,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)
7.7 栈解码 223
null
0
P =1
on
table
there is
2
1
3
P =0.2
P =0.3
P=0.5
one
an apple
table
on the table
apple
4
4-5
1
1-2
5
P =0.1
P =0.4
P =0.3
P =0.4
P =0.2
翻译路径:
图 7.25 翻译假设扩展
7.7.3 剪枝
假设扩展建立了解码算法的基本框架。但是,当句子变长时,这种方法还是面
临着搜索空间爆炸的问题。对于这个问题,常用的解决办法是剪枝,也就是在搜索
图中排除掉一些节点。比如,可以使用束剪枝,确保每次翻译扩展时,最多生成 k 个
新的翻译假设。这里 k 可以被看做是束的宽度。通过控制 k 的大小,可以在解码精
度和速度之间进行平衡。这种基于束宽度进行剪枝的方法也被称作直方图剪枝。另
一种思路是,每次扩展时只保留与最优翻译假设得分相差在 δ 之内的翻译假设。δ 可
以被看作是一种与最优翻译假设之间距离的阈值,超过这个阈值就被剪枝。这种方
法也被称作阈值剪枝(Threshold Pruning)。
不过,即使引入束剪枝,解码过程中仍然会有很多冗余的翻译假设。有两种方
法可以进一步加速解码:
• 对相同译文的翻译假设进行重新组合;
• 对低质量的翻译假设进行裁剪。
对翻译假设进行重新组合又被称作假设重组(Hypothesis Recombination)。其核
心思想是,把代表同一个译文的不同翻译假设融合为一个翻译假设。如图7.26所示,
对于给定的输入短语“一个 苹果”,系统可能将两个单词“一个”、“苹果”分别翻
译成“an”和“apple”,也可能将这两个单词作为一个短语直接翻译成“an apple”。
虽然这两个翻译假设得到的译文相同,并且覆盖了相同的源语言短语,但是却是两
个不同的翻译假设,模型给它们的打分也是不一样的。这时,可以舍弃两个翻译假
设中分数较低的那个,因为分数较低的翻译假设永远不可能成为最优路径的一部分。
这也就相当于把两个翻译假设重组为一个假设。
即使翻译假设对应的译文不同也可以进行假设重组。图7.26的下半部分给出了
一个这样的实例。在两个翻译假设中,第一个单词分别被翻译成了“it”和“he”,紧
接着它们后面的部分都被翻译成了“is not”。这两个翻译假设是非常相似的,因为它
们译文的最后两个单词是相同的,而且翻译假设都覆盖了相同的源语言部分。这时,
也可以对这两个翻译假设进行假设重组:如果得分较低的翻译假设和得分较高的翻