BLEU

BLEU (Bilingual Evaluation Understudy) is an algorithm for evaluating the quality of text which has been machine-translated from one natural language to another. Quality is considered to be the correspondence between a machine's output and that of a human: "the closer a machine translation is to a professional human translation, the better it is".[1] BLEU was one of the first metrics to achieve a high correlation with human judgements of quality,[2][3] and remains one of the most popular.

Scores are calculated for individual translated segments—generally sentences—by comparing them with a set of good quality reference translations. Those scores are then averaged over the whole corpus to reach an estimate of the translation's overall quality. Intelligibility or grammatical correctness are not taken into account.

BLEU is designed to approximate human judgement at a corpus level, and performs badly if used to evaluate the quality of individual sentences.

BLEU’s output is always a number between 0 and 1. This value indicates how similar the candidate and reference texts are, with values closer to 1 representing more similar texts.

Algorithm

BLEU uses a modified form of precision to compare a candidate translation against multiple reference translations. The metric modifies simple precision since machine translation systems have been known to generate more words than appear in a reference text. This is illustrated in the following example from Papineni et al. (2002),

Example of poor machine translation output with high precision
Candidate the the the the the the the
Reference 1 the cat is on the mat
Reference 2 there is a cat on the mat

Of the seven words in the candidate translation, all of them appear in the reference translations. Thus the candidate text is given a unigram precision of,

P = \frac{m}{w_{t}} = \frac{7}{7} = 1

where ~m is number of words from the candidate that are found in the reference, and ~w_{t} is the total number of words in the candidate. This is a perfect score, despite the fact that the candidate translation above retains little of the content of either of the references.

The modification that BLEU makes is fairly straightforward. For each word in the candidate translation, the algorithm takes its maximum total count, ~m_{max}, in any of the reference translations. In the example above, the word "the" appears twice in reference 1, and once in reference 2. Thus ~m_{max} = 2.

For the candidate translation, the count m_{w} of each word is clipped to a maximum of m_{max} for that word. In this case, "the" has ~m_{w} = 7 and ~m_{max}=2, thus ~m_{w} is clipped to 2. ~m_{w} is then summed over all words in the candidate. This sum is then divided by the total number of words in the candidate translation. In the above example, the modified unigram precision score would be:

P = \frac{2}{7}

The above method is used to calculate scores for a range of n-gram lengths. The length which has the "highest correlation with monolingual human judgements"[4] was found to be four. The unigram scores are found to account for the adequacy of the translation, how much information is retained. The longer n-gram scores account for the fluency of the translation, or to what extent it reads like "good English".

The modification made to precision does not solve the problem of short translations, which can produce very high precision scores, even using modified precision. An example of a candidate translation for the same references as above might be:

the cat

In this example, the modified unigram precision would be,

P = \frac{1}{2} %2B \frac{1}{2} = \frac{2}{2}

as the word 'the' and the word 'cat' appear once each in the candidate, and the total number of words is two. The modified bigram precision would be 1 / 1 as the bigram, "the cat" appears once in the candidate. It has been pointed out that precision is usually twinned with recall to overcome this problem [5], as the unigram recall of this example would be 2 / 6 or 2 / 7. The problem being that as there are multiple reference translations, a bad translation could easily have an inflated recall, such as a translation which consisted of all the words in each of the references.[6]

In order to produce a score for the whole corpus the modified precision scores for the segments are combined, using the geometric mean multiplied by a brevity penalty to prevent very short candidates from receiving too high a score. Let r be the total length of the reference corpus, and c the total length of the translation corpus. If c \leq r, the brevity penalty applies, defined to be e^{(1-r/c)}. (In the case of multiple reference sentences, r is taken to be the sum of the lengths of the sentences whose lengths are closest to the lengths of the candidate sentences. However, in the version of the metric used by NIST evaluations prior to 2009, the shortest reference sentence had been used instead.)

iBLEU is an interactive version of BLEU that allows a user to visually examine the BLEU scores obtained by the candidate translations. It also allows comparing two different systems in a visual and interactive manner which is useful for system development.[7]

Performance

BLEU has frequently been reported as correlating well with human judgement,[8][9][10] and remains a benchmark for the assessment of any new evaluation metric. There are however a number of criticisms that have been voiced. It has been noted that although in principle capable of evaluating translations of any language, BLEU cannot in its present form deal with languages lacking word boundaries.[11]

It has been argued that although BLEU has significant advantages, there is no guarantee that an increase in BLEU score is an indicator of improved translation quality.[12] Nevertheless, they highlight two instances where BLEU seriously underperformed. These were the 2005 NIST evaluations[13] where a number of different machine translation systems were tested, and their study of the SYSTRAN engine versus two engines using statistical machine translation (SMT) techniques.[14]

In the 2005 NIST MT evaluation, it is reported that the scores generated by BLEU failed to correspond to the scores produced in the human evaluations. The system which was ranked highest by the human judges was only ranked 6th by BLEU. In their study, they compared SMT systems with SYSTRAN, a knowledge based system. The scores from BLEU for SYSTRAN were substantially worse than the scores given to SYSTRAN by the human judges. They note that the SMT systems were trained using BLEU minimum error rate training,[15] and point out that this could be one of the reasons behind the difference. They conclude by recommending that BLEU be used in a more restricted manner, for comparing the results from two similar systems, and for tracking "broad, incremental changes to a single system".[16]

See also

Notes

  1. ^ Papineni, K., et al. (2002)
  2. ^ Papineni, K., et al. (2002)
  3. ^ Coughlin, D. (2003)
  4. ^ Papineni, K., et al. (2002)
  5. ^ Papineni, K., et al. (2002)
  6. ^ Papineni, K., et al. (2002)
  7. ^ Papineni, K., et al. (2002)
  8. ^ Coughlin, D. (2003)
  9. ^ Doddington, G. (2002)
  10. ^ Denoual, E. and Lepage, Y. (2005)
  11. ^ Callison-Burch, C., Osborne, M. and Koehn, P. (2006)
  12. ^ Lee, A. and Przybocki, M. (2005)
  13. ^ Callison-Burch, C., Osborne, M. and Koehn, P. (2006)
  14. ^ Lin, C. and Och, F. (2004)
  15. ^ Callison-Burch, C., Osborne, M. and Koehn, P. (2006)
  16. ^ Madnani, N. (2011)

References