一、神经机器翻译(Neural Machine Translation)评价指标之BLEU详细计算过程
1. 简介
BLEU(Bilingual Evaluation Understudy),相信大家对这个评价指标的概念已经很熟悉,随便百度谷歌就有相关介绍。原论文为BLEU: a Method for Automatic Evaluation of Machine Translation,IBM出品。
本文通过一个例子详细介绍BLEU是如何计算以及NLTKnltk.align.bleu_score模块的源码。
首先祭出公式:
注意这里的BLEU值是针对一条翻译(一个样本)来说的。
NLTKnltk.align.bleu_score模块实现了这里的公式,主要包括三个函数,两个私有函数分别计算P和BP,一个函数整合计算BLEU值。
# 计算BLEU值
def bleu(candidate, references, weights)
# (1)私有函数,计算修正的n元精确率(Modified n-gram Precision)
def _modified_precision(candidate, references, n)
# (2)私有函数,计算BP惩罚因子
def _brevity_penalty(candidate, references)
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
例子:
候选译文(Predicted):
It is a guide to action which ensures that the military always obeys the commands of the party
参考译文(Gold Standard)
1:It is a guide to action that ensures that the military will forever heed Party commands
2:It is the guiding principle which guarantees the military forces always being under the command of the Party
3:It is the practical guide for the army always to heed the directions of the party
2. Modified n-gram Precision计算(也即是PnPn)
def _modified_precision(candidate, references, n):
counts = Counter(ngrams(candidate, n))
if not counts:
return 0
max_counts = {}
for reference in references:
reference_counts = Counter(ngrams(reference, n))
for ngram in counts:
max_counts[ngram] = max(max_counts.get(ngram, 0), reference_counts[ngram])
clipped_counts = dict((ngram, min(count, max_counts[ngram])) for ngram, count in counts.items())
return sum(clipped_counts.values()) / sum(counts.values())
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
我们这里nn取值为4,也就是从1-gram计算到4-gram。
Modified 1-gram precision:
首先统计候选译文里每个词出现的次数,然后统计每个词在参考译文中出现的次数,Max表示3个参考译文中的最大值,Min表示候选译文和Max两个的最小值。
词 | 候选译文 | 参考译文1 | 参考译文2 | 参考译文3 | Max | Min |
---|---|---|---|---|---|---|
the | 3 | 1 | 4 | 4 | 4 | 3 |
obeys | 1 | 0 | 0 | 0 | 0 | 0 |
a | 1 | 1 | 0 | 0 | 1 | 1 |
which | 1 | 0 | 1 | 0 | 1 | 1 |
ensures | 1 | 1 | 0 | 0 | 1 | 1 |
guide | 1 | 1 | 0 | 1 | 1 | 1 |
always | 1 | 0 | 1 | 1 | 1 | 1 |
is | 1 | 1 | 1 | 1 | 1 | 1 |
of | 1 | 0 | 1 | 1 | 1 | 1 |
to | 1 | 1 | 0 | 1 | 1 | 1 |
commands | 1 | 1 | 0 | 0 | 1 | 1 |
that | 1 | 2 | 0 | 0 | 2 | 1 |
It | 1 | 1 | 1 | 1 | 1 | 1 |
action | 1 | 1 | 0 | 0 | 1 | 1 |
party | 1 | 0 | 0 | 1 | 1 | 1 |
military | 1 | 1 | 1 | 0 | 1 | 1 |
然后将每个词的Min值相加,将候选译文每个词出现的次数相加,然后两值相除即得P1=3+0+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+13+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1=0.95P1=3+0+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+13+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1=0.95。
类似可得:
Modified 2-gram precision:
词 | 候选译文 | 参考译文1 | 参考译文2 | 参考译文3 | Max | Min |
---|---|---|---|---|---|---|
ensures that | 1 | 1 | 0 | 0 | 1 | 1 |
guide to | 1 | 1 | 0 | 0 | 1 | 1 |
which ensures | 1 | 0 | 0 | 0 | 0 | 0 |
obeys the | 1 | 0 | 0 | 0 | 0 | 0 |
commands of | 1 | 0 | 0 | 0 | 0 | 0 |
that the | 1 | 1 | 0 | 0 | 1 | 1 |
a guide | 1 | 1 | 0 | 0 | 1 | 1 |
of the | 1 | 0 | 1 | 1 | 1 | 1 |
always obeys | 1 | 0 | 0 | 0 | 0 | 0 |
the commands | 1 | 0 | 0 | 0 | 0 | 0 |
to action | 1 | 1 | 0 | 0 | 1 | 1 |
the party | 1 | 0 | 0 | 1 | 1 | 1 |
is a | 1 | 1 | 0 | 0 | 1 | 1 |
action which | 1 | 0 | 0 | 0 | 0 | 0 |
It is | 1 | 1 | 1 | 1 | 1 | 1 |
military always | 1 | 0 | 0 | 0 | 0 | 0 |
the military | 1 | 1 | 1 | 0 | 1 | 1 |
P2=1017=0.588235294P2=1017=0.588235294
Modified 3-gram precision:
词 | 候选译文 | 参考译文1 | 参考译文2 | 参考译文3 | Max | Min |
---|---|---|---|---|---|---|
ensures that the | 1 | 1 | 0 | 0 | 1 | 1 |
which ensures that | 1 | 0 | 0 | 0 | 0 | 0 |
action which ensures | 1 | 0 | 0 | 0 | 0 | 0 |
a guide to | 1 | 1 | 0 | 0 | 1 | 1 |
military always obeys | 1 | 0 | 0 | 0 | 0 | 0 |
the commands of | 1 | 0 | 0 | 0 | 0 | 0 |
commands of the | 1 | 0 | 0 | 0 | 0 | 0 |
to action which | 1 | 0 | 0 | 0 | 0 | 0 |
the military always | 1 | 0 | 0 | 0 | 0 | 0 |
obeys the commands | 1 | 0 | 0 | 0 | 0 | 0 |
It is a | 1 | 1 | 0 | 0 | 1 | 1 |
of the party | 1 | 0 | 0 | 1 | 1 | 1 |
is a guide | 1 | 1 | 0 | 0 | 1 | 1 |
that the military | 1 | 1 | 0 | 0 | 1 | 1 |
always obeys the | 1 | 0 | 0 | 0 | 0 | 0 |
guide to action | 1 | 1 | 0 | 0 | 1 | 1 |
P3=716=0.4375P3=716=0.4375
Modified 4-gram precision:
词 | 候选译文 | 参考译文1 | 参考译文2 | 参考译文3 | Max | Min |
---|---|---|---|---|---|---|
to action which ensures | 1 | 0 | 0 | 0 | 0 | 0 |
action which ensures that | 1 | 0 | 0 | 0 | 0 | 0 |
guide to action which | 1 | 0 | 0 | 0 | 0 | 0 |
obeys the commands of | 1 | 0 | 0 | 0 | 0 | 0 |
which ensures that the | 1 | 0 | 0 | 0 | 0 | 0 |
commands of the party | 1 | 0 | 0 | 0 | 0 | 0 |
ensures that the military | 1 | 1 | 0 | 0 | 1 | 1 |
a guide to action | 1 | 1 | 0 | 0 | 1 | 1 |
always obeys the commands | 1 | 0 | 0 | 0 | 0 | 0 |
that the military always | 1 | 0 | 0 | 0 | 0 | 0 |
the commands of the | 1 | 0 | 0 | 0 | 0 | 0 |
the military always obeys | 1 | 0 | 0 | 0 | 0 | 0 |
military always obeys the | 1 | 0 | 0 | 0 | 0 | 0 |
is a guide to | 1 | 1 | 0 | 0 | 1 | 1 |
It is a guide | 1 | 1 | 0 | 0 | 1 | 1 |
P4=415=0.266666667P4=415=0.266666667
然后我们取w1=w2=w3=w4=0.25w1=w2=w3=w4=0.25,也就是Uniform Weights。
所以:
∑Ni=1wnlogPn=0.25∗logP1+0.25∗logP2+0.25∗logP3+0.25∗logP4=−0.684055269517∑i=1NwnlogPn=0.25∗logP1+0.25∗logP2+0.25∗logP3+0.25∗logP4=−0.684055269517
3. Brevity Penalty 计算
def _brevity_penalty(candidate, references):
c = len(candidate)
ref_lens = (len(reference) for reference in references)
#这里有个知识点是Python中元组是可以比较的,如(0,1)>(1,0)返回False,这里利用元组比较实现了选取参考翻译中长度最接近候选翻译的句子,当最接近的参考翻译有多个时,选取最短的。例如候选翻译长度是10,两个参考翻译长度分别为9和11,则r=9.
r = min(ref_lens, key=lambda ref_len: (abs(ref_len - c), ref_len))
print 'r:',r
if c > r:
return 1
else:
return math.exp(1 - r / c)
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
下面计算BP(Brevity Penalty),翻译过来就是“过短惩罚”。由BP的公式可知取值范围是(0,1],候选句子越短,越接近0。
候选翻译句子长度为18,参考翻译分别为:16,18,16。
所以c=18c=18)
所以BP=e0=1BP=e0=1
4. 整合
最终BLEU=1⋅exp(−0.684055269517)=0.504566684006BLEU=1⋅exp(−0.684055269517)=0.504566684006。
BLEU的取值范围是[0,1],0最差,1最好。
通过计算过程,我们可以看到,BLEU值其实也就是“改进版的n-gram”加上“过短惩罚因子”。
二、注:
BLEU score 用来评价一个翻译系统的好坏。计算公式如下:
N指使用N-gram计算Pn。BLEU越高越好。
1. 计算Pn
Pn = (候选译文与参考译文相同的N-gram数目) / (候选译文中所有N-gram的数目)
wn为权重,可以取平均值1/N
当有多个标准答案时,取min(max(参考译文),候选译文)。
2. 计算惩罚项BP
c指候选译文(预测译文)的长度,r指(所有参考译文中,其长度与候选译文最接近的长度)
相关文章:https://blog.csdn.net/guolindonggld/article/details/56966200
3. BLEU score现状
WMT’14 English to French: 34左右
参考:
- 机器翻译评价指标之BLEU详细计算过程
- 【NLP】BLEU
- 【NLP】机器翻译常用评价标准 (BLEU & METEOR)
- 自然语言处理——BLEU详解以及简单的代码实现
- 一种机器翻译的评价准则——Bleu
- 【NLP】文本生成评价指标:BLEU
- 机器翻译评测——BLEU算法详解
- 机器翻译自动评估-BLEU算法详解
附: cs224n-2019-winter 作业四 写作部分 计算BLEU分数