CHRF评估指标

CHRF

  1. CHRF指标从字符级别对译文质量进行评估,它考虑了一些形态—句法现象,除此之外,与其他评估指标相比,它很简单,不需要任何额外的工具或知识来源,它完全独立于语言,也独立于分词过程。

  2. CHRF计算公式:
    c h r F β = ( 1 + β 2 ) c h r P ⋅ c h r R β 2 ⋅ c h r P + c h r R \mathrm{chrF} \beta=\left(1+\beta^{2}\right) \frac{\mathrm{chrP} \cdot \mathrm{chrR}}{\beta^{2} \cdot \mathrm{chrP}+\mathrm{chrR}} chrFβ=(1+β2)β2chrP+chrRchrPchrR

    • c h r P \mathrm{chrP} chrP是精确度,指翻译句子和参考译文句子匹配的字符级n-gram在翻译句子中占的比例
    • c h r R \mathrm{chrR} chrR是召回率,指翻译句子和参考译文句子匹配的字符级n-gram在参考译文句子中占的比例
    • β \beta β可以控制召回率和精确度两个指标的重要性(召回率比准确率重要 β \beta β倍),当 β = 1 \beta=1 β=1时二者同样重要
  3. 使用nltk计算CHRF

    • n-gram词组长度为1时(词组的最小长度为1,最大长度也为1, β = 3 \beta=3 β=3

      from nltk.translate.chrf_score import sentence_chrf
      
      ref = 'the cat is on the mat'.split()
      hyp = 'the the the the the the the'.split()
      sentence_chrf(ref, hyp, min_len=1, max_len=1, beta=3.0)
      # 0.48484848484848486
      
      • 计算TP:重复出现的1-gram{t, h, e},总共有8次

      • 计算TP + FP:翻译句子的长度为21

      • 计算TP + TN:参考译文句子的长度为16

      • c h r P = 8 / 21 , c h r R = 8 / 16 \mathrm{chrP}=8/21,\mathrm{chrR}=8/16 chrP=8/21,chrR=8/16
        c h r F = ( 1 + 3 2 ) 8 21 ∗ 1 2 3 2 ∗ 8 21 + 1 2 = 16 33 = 0.48484848484848485 \mathrm{chrF}=(1 + 3^2)\frac{\frac{8}{21}*\frac{1}{2}}{3^2*\frac{8}{21}+\frac{1}{2}}=\frac{16}{33}=0.48484848484848485 chrF=(1+32)32218+2121821=3316=0.48484848484848485

    • 拓展到max_len=2的情况,此时n-gram词组的最小长度为1,最大长度为2, β = 3 \beta=3 β=3

      from nltk.translate.chrf_score import sentence_chrf
      ref = 'the cat is on the mat'.split()
      hyp = 'the the the the the the the'.split()
      print(sentence_chrf(ref, hyp, min_len=1, max_len=2, beta=3.0))
      # 0.37145650048875856
      
      • 计算1-gram的情况:此时F-score=0.48484848484848486(和CHRF计算相同)

      • 计算2-gram的情况:

        • 计算TP:重复出现的2-gram{th, he},总共有4次

        • 计算TP + FP: 翻译句子分成2-gram的长度为20

        • 计算TP + TN:参考译文句子分成2-gram的长度为15

        • c h r P = 4 / 20 , c h r R = 4 / 15 \mathrm{chrP}=4/20,\mathrm{chrR}=4/15 chrP=4/20,chrR=4/15
          c h r F 2 − g r a m = ( 1 + 3 2 ) 1 5 ∗ 4 15 3 2 ∗ 1 5 + 4 15 = 8 31 = 0.25806451612903225 \mathrm{chrF_{2-gram}}=(1 + 3^2)\frac{\frac{1}{5}*\frac{4}{15}}{3^2*\frac{1}{5}+\frac{4}{15}}=\frac{8}{31}=0.25806451612903225 chrF2gram=(1+32)3251+15451154=318=0.25806451612903225

      • 计算总的CHRF
        c h r F = 0.48484848484848486 + 0.25806451612903225 2 = 0.37145650048875856 \mathrm{chrF}=\frac{0.48484848484848486+0.25806451612903225}{2}=0.37145650048875856 chrF=20.48484848484848486+0.25806451612903225=0.37145650048875856

    • 计算语料级的CHRF(以上都是句子级的CHRF):基本思想是计算出每个句子的CHRF,然后再求算术平均

      ref1 = str('It is a guide to action that ensures that the military will forever heed Party commands').split()
      ref2 = str('It is the guiding principle which guarantees the military forces always being under the command of the Party').split()
      hyp1 = str('It is a guide to action which ensures that the military always obeys the commands of the party').split()
      hyp2 = str('It is to insure the troops forever hearing the activity guidebook that party direct').split()
      corpus_chrf([ref1, ref2], [hyp1, hyp2]) 
      # 0.4166529443281564
      
      (sentence_chrf(ref1, hyp1) + sentence_chrf(ref2, hyp2)) / 2
      # 0.4166529443281564
      
  4. 使用sacrebleu计算CHRF

    • 计算句子级CHRF

      print(sacrebleu.sentence_chrf(hypothesis='the the the the the the the',
                                  references=['the cat is on the mat'],
                                  char_order=1, word_order=0, beta=3, remove_whitespace=True).score)
      # 48.484848484848484
      print(sacrebleu.sentence_chrf(hypothesis='the the the the the the the',
                                  references=['the cat is on the mat'],
                                  char_order=2, word_order=0, beta=3, remove_whitespace=True).score)
      # 37.145882975906794
      
    • 计算语料级CHRF

      • nltk工具提供的计算方法不同,sacrebleu并不是计算出每个句子的CHRF,再求算术平均
      • sacrebleu在计算i-gram的准确率和召回率时,将语料中的参考句子i-gram长度、翻译句子i-gram长度、参考句子和翻译句子匹配i-gram数量分别进行相加,即分数中的分子和分子进行相加,分母和分母进行相加,与nltk中的分数直接进行相加不同,这与sacrebleu中求BLEU的方法有异曲同工之妙
      ref1 = 'It is a guide to action that ensures that the military will forever heed Party commands'
      ref2 = 'It is the guiding principle which guarantees the military forces always being under the command of the Party'
      hyp1 = 'It is a guide to action which ensures that the military always obeys the commands of the party'
      hyp2 = 'It is to insure the troops forever hearing the activity guidebook that party direct'
      print(sacrebleu.corpus_chrf(hypotheses=[hyp1, hyp2], references=[[ref1, ref2]], char_order=6, word_order=0, beta=3).score)
      # 39.364938843711016
      
      • 将以上的代码作为示例,sacrebleu首先计算出每个句子各n-gram模型中的参考句子n-gram长度、翻译句子n-gram长度、参考句子和翻译句子匹配n-gram数量,如下所示

        [[77, 72, 65, 76, 71, 50, 75, 70, 44, 74, 69, 40, 73, 68, 36, 72, 67, 33], [70, 91, 60, 69, 90, 28, 68, 89, 12, 67, 88, 4, 66, 87, 1, 65, 86, 0]]
        
      • 列表中的第一项代表第一个句子,列表中的第二项代表第二个句子,以第一个句子为例,列表项中共有18个元素,分别是翻译句子1-gram长度(翻译句子长度)、参考句子1-gram长度(参考句子长度)、翻译句子和参考句子匹配的1-gram数量…以此类推,一直到6-gram

      • 将列表中的对应项进行相加,得出以下结果(相当于分子和分子相加,分母和分母相加)

        [147, 163, 125, 145, 161, 78, 143, 159, 56, 141, 157, 44, 139, 155, 37, 137, 153, 33]
        
      • 然后求各n-gram的准确率和召回率,将准确率和召回率求算术平均(总和除以6),再用平均后的准确率和召回率求最终的CHRF,基础逻辑如下所示(仿照sacrebleu手搓的,可能有一些特殊情况不适用,比如分母不能为0)

        data_list = [147, 163, 125, 145, 161, 78, 143, 159, 56, 141, 157, 44, 139, 155, 37, 137, 153, 33]
        sum_prec, sum_rec = 0, 0
        for i in range(0, 6):
            index = 3 * i
            n_hyp = data_list[index]
            n_ref = data_list[index + 1]
            n_match = data_list[index + 2]
            
            n_prec = n_match / n_hyp
            n_rec = n_match / n_ref
            
            sum_prec += n_prec
            sum_rec += n_rec
            
            n_fscore = (1 + 9) * n_prec * n_rec / (9 * n_prec + n_rec)
            sum_fscore = sum_fscore + n_fscore
        
        print((1 + 9) * (sum_prec / 6) * (sum_rec / 6) / (9 * (sum_prec / 6) + (sum_rec / 6)))
        # 0.39364938843711017
        

CHRF++

  1. 之前的工作中显示,对于评分较差的句子,CHRF和WORDF分数的标准差是相似的——两个指标都分配了相对相似的(低)分数,但对于人类评分较高的句子,CHRF的偏差相较于WORDF的偏差要低得多,此外,人类评分越高,WORDF与CHRF的偏差的差异越大,这些结果表明,CHRF是优于WORDF的,尤其是在翻译质量较高的片段上

  2. 但是考虑到CHRF的结果可能过于乐观,所以将CHRF和WORDF结合起来,得到CHRF++

  3. 当单词n-grams与字符n-grams相加并取平均值时,就会得到CHRF++分数,这种组合的最佳n-gram长度对于字符n-gram来说是n=6,与CHRF中字符n-gram的最佳长度相同,对于单词n-gram来说是n=1n=2

  4. 使用sacrebleu计算CHRF++

    print(sacrebleu.sentence_chrf(hypothesis='the the the the the the the',
                                      references=['the cat is on the mat'],
                                      char_order=1, word_order=1, beta=3, remove_whitespace=True).score)
    # 40.65040650406503
    
    • 为了方便计算,这里字符和单词都选择1-gram,首先可以得到以下统计结果

      [21, 16, 8, 7, 6, 2]
      
    • 前三个数是字符级1-gram的统计结果,后三个数是单词级1-gram的统计结果(分别是翻译句子字符或单词1-gram长度、参考句子字符或单词1-gram长度、匹配的1-gram数量)

    • 分别计算准确率和召回率并求平均值(实际上sacrebleu中求平均值是除以self.order实现的,此时self.order等于列表长度除以3,本例中为 6 / 3 = 2 6/3=2 6/3=2),得到最后的准确率和召回率,再计算CHRF++

      prec = 0.6666666666666666 / 2
      rec = 0.8333333333333333 / 2
      (1 + 9) * prec * rec / (9 * prec + rec)
      # 0.40650406504065034
      

参考文献:

  1. NLTK :: nltk.translate.chrf_score
  2. NLG评估指标chrF、chrF++介绍_地大陈参志的博客-CSDN博客
  3. sacrebleu/chrf.py at master · mjpost/sacrebleu (github.com)
  4. chrF: character n-gram F-score for automatic MT evaluation - ACL Anthology
  5. chrF++: words helping character n-grams - ACL Anthology
  • 2
    点赞
  • 4
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值