BLEU {{TOCRight}} :''This page is about the evaluation metric for machine translation. For other meanings, please see [[Bleu]].'' '''BLEU''' ('''Bilingual Evaluation Understudy''') is a method for evaluating the quality of text which has been translated from one [[natural language]] to another using [[machine translation]]. BLEU was one of the first [[software metric]]s to report high [[correlation]] with human judgements of quality. The metric is currently one of the most popular in the field. The central idea behind the metric is that, "the closer a machine translation is to a professional human translation, the better it is".{{ref|Papineni2002a}} The metric calculates scores for individual segments, generally [[Sentence (linguistics)|sentence]]s, and then averages these scores over the whole [[corpus]] in order to reach a final score. It has been shown to correlate highly with human judgements of quality at the corpus level.{{ref|Papineni2002b}}{{ref|Coughlin2003a}} The quality of translation is indicated as a number between 0 and 1 and is measured as statistical closeness to a given set of good quality human reference translations. Therefore, it does not directly take into account translation intelligibility or grammatical correctness. The metric works by measuring the [[n-gram]] co-occurrence between a given translation and the set of reference translations and then taking the weighted [[geometric mean]]. BLEU is specifically designed to approximate human judgement on a [[corpus]] level and performs badly if used to evaluate the quality of isolated sentences. ==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), {|class=wikitable |+ 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 |- |} In this example, the candidate text is given a unigram precision of, :P = \frac{m}{w_{t}} = \frac{7}{7} = 1 Of the seven words in the candidate translation, all of them appear in the reference translations. This presents a problem for a metric, as the candidate translation above is complete nonsense, retaining none 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 the maximum total count in the reference translations. Taking the example above, the word 'the' appears twice in reference 1, and once in reference 2. The largest value is taken, in this case '2' as the "maximum reference count". For each of the words in the candidate translation, the count of the word is compared against the maximum reference count, and the lowest value is taken. In this case, the count of the word 'the' in the candidate translation is '7', while the maximum reference count for the word is '2'. This "modified count" 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 each n. The value of n which has the "highest correlation with monolingual human judgements"{{ref|Papineni2002c}} was found to be 4. The unigram scores are found to account for the adequacy of the translation, in other words, how much information is retained in the translation. 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. Short translations 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} + \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 {{ref|Papineni2002d}}, 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.{{ref|Papineni2002e}} 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, whose purpose is 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 and is 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]], the short reference sentence is used.) ==Performance== BLEU has frequently been reported as correlating well with human judgement,{{ref|Papineni2002f}}{{ref|Coughlin2003b}}{{ref|Doddington2002a}} and certainly remains a benchmark for any new evaluation metric to beat. There are however a number of criticisms that have been voiced. It has been noted that while in theory capable of evaluating any language, BLEU does not in the present form work on languages without word boundaries.{{ref|Denoul2005a}} It has been argued that although BLEU certainly has significant advantages, there is no guarantee that an increase in BLEU score is an indicator of improved translation quality.{{ref|Callison2006a}} As BLEU scores are taken at the corpus level, it is difficult to give a textual example. Nevertheless, they highlight two instances where BLEU seriously underperformed. These were the 2005 [[NIST]] evaluations{{ref|Lee2005a}} 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.{{ref|Callison2006b}} In the 2005 NIST evaluation, they report 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,{{ref|Och2004a}} 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".{{ref|Callison2006c}} ==See also== * [[NIST (metric)]] * [[METEOR]] ==Notes==
# {{note|Papineni2002a}} Papineni, K., et al. (2002) # {{note|Papineni2002b}} Papineni, K., et al. (2002) # {{note|Coughlin2003a}} Coughlin, D. (2003) # {{note|Papineni2002c}} Papineni, K., et al. (2002) # {{note|Papineni2002d}} Papineni, K., et al. (2002) # {{note|Papineni2002e}} Papineni, K., et al. (2002) # {{note|Papineni2002e}} Papineni, K., et al. (2002) # {{note|Coughlin2003b}} Coughlin, D. (2003) # {{note|Doddington2002a}} Doddington, G. (2002) # {{note|Denoul2005a}} Denoul, E. and Lepage, Y. (2005) # {{note|Callison2006a}} Callison-Burch, C., Osborne, M. and Koehn, P. (2006) # {{note|Lee2005a}} Lee, A. and Przybocki, M. (2005) # {{note|Callison2006b}} Callison-Burch, C., Osborne, M. and Koehn, P. (2006) # {{note|Och2004a}} Lin, C. and Och, F. (2004) # {{note|Callison2006c}} Callison-Burch, C., Osborne, M. and Koehn, P. (2006)
==References== * Papineni, K., Roukos, S., Ward, T., and Zhu, W. J. (2002). "BLEU: a method for automatic evaluation of machine translation" in ''ACL-2002: 40th Annual meeting of the Association for Computational Linguistics'' pp. 311--318 * Callison-Burch, C., Osborne, M. and Koehn, P. (2006) "Re-evaluating the Role of BLEU in Machine Translation Research" in ''11th Conference of the European Chapter of the Association for Computational Linguistics: EACL 2006'' pp. 249--256 * Doddington, G. (2002) "Automatic evaluation of machine translation quality using n-gram cooccurrence statistics" in ''Proceedings of the Human Language Technology Conference (HLT), San Diego, CA'' pp. 128--132 * Coughlin, D. (2003) "Correlating Automated and Human Assessments of Machine Translation Quality" in ''MT Summit IX, New Orleans, USA'' pp. 23--27 * Denoul, E. and Lepage, Y. (2005) "BLEU in characters: towards automatic MT evaluation in languages without word delimiters" in ''Companion Volume to the Proceedings of the Second International Joint Conference on Natural Language Processing'' pp. 81--86 * Lee, A. and Przybocki, M. (2005) NIST 2005 machine translation evaluation official results * Lin, C. and Och, F. (2004) "Automatic Evaluation of Machine Translation Quality Using Longest Common Subsequence and Skip-Bigram Statistics" in ''Proceedings of the 42nd Annual Meeting of the Association of Computational Linguistics''. ==External links== [[Category:Evaluation of machine translation]] [[es:BLEU]] [[fa:جایگزین ارزشیابی دوزبانه]] [[tg:BLEU]]