Orange: a Method for Evaluating Automatic Evaluation Metrics for Machine Translation Chin-Yew Lin & Franz Josef Och (presented by Bilmes) or Orange: a Method for Automatically Evaluating Automatic Evaluation Metrics for Machine Translation
Summary 1. Introduces ORANGE, a way to automatically evaluate automatic evaluation methods 2. Introduces 3 new ways to automatically evaluate MT systems: ROUGE-L, ROUGE-W, and ROUGE-S 3. Uses ORANGE to evaluate many different evaluation methods, and finds that their new one, ROUGE-S4 is the best evaluator
Reminder: Adequacy & Fluency Adequacy refers to the degree to which the translation communicates information present in the original. Roughly, a translation using the same words (1-grams) as the reference tends to satisfy adequecy. Fluency refers to the degree to which the translation is well-formed according to the grammar of the target language. Roughly, the longer the n-gram matches of a translation with a reference, tends to improve fluency.
Reminder: BLEU unigram precision = num translation unigrams that appear in reference translation/candidate translation length modified unigram precision = clipped(num trans. unigrams in ref translation)/cand. trans. length –clipping maxes out to the max count in any ref. translation modified n-gram precision (same thing) On blocks of text: Brevity penalty (since short sentences can get high low- gram precision) Finally:
Still Other Reminders Person’s product moment correlation coefficient: –just the normal correlation coefficient r 2 = (EXY) 2 /EX 2 EY 2 Spearman’s rank order correlation coefficient –same thing for normal orders, or otherwise: –D i = rank A (i) – rank B (i) Bootstrap method to compute confidence intervals –resample with replacement from data N times, compute mean and get confidence interval of val +/- 2se(val) for 95% confidence interval.
On to the paper: Lots of ways to evaluate MT quality BLEU RED WER – length-normalized edit distance PER – position independent word error rate (bag of words approach) GTM – general text matcher, based on a balance of recall, precision, & their F-measure combination (we should do this paper) This paper now introduces still three more such metrics: ROUGE-L, ROUGE-W, and ROUGE-S (which we shall define).
Corr Coeff & 95% CIs of 8 MT systems in NIST03 Chinese-English, using various MT evaluation methods
Problem is we need a way to automatically evaluate these automatic evaluation methods. Since we don’t know which one is best, which one to use, how and when to choose, etc. Try to break out of the region of insignificant difference. Question: what about meta regress: do we need a way to automatically evaluate automatic evaluations of automatic evaluation methods? Anyway, goal of this paper (other than introducing new automatic evaluation methods) is to introduce ORANGE: Oracle Ranking for Gisting Evaluation (or the first automatic evaluation of automatic MT evaluation methods).
ORANGE Intuitively: uses translations “rank” as scored by MT evaluation system (so good translations should have high rank, poor ones should have low rank) reference translations should have higher rank. Key quantity: average rank of reference translations within combined machine and reference translation list. ORANGE = average rank / N in N-best list. 1.The bank was visited by me yesterday. 2.I went to the bank yesterday 3.Yesterday, I went to the bank. 4.Yesterday, the bank had the opportunity to be visited by me, and in fact this did indeed occur. 5.There was once this back that at least as of yesterday existed, and so did I, and a funny thing happened … So, reference translations were ranked 2 and 3 in this list. Avg rank = 2.5. Smaller the better.
ORANGE The way they calculate ORANGE in this work: Oracle i = reference transcription I N = size of N-best list S = number of sentences in corpus Rank(Oracle i ) = average rank of source sentence i’s reference translations in n-best list i.
Three new metrics ROUGE-L ROUGE-W ROUGE-S
LCS: Longest Common Subsequences
Computing: Longest Common Subsequences 1 Key thing: This does not require consecutive matches in strings. Ex: LCS(X,Y) = 3 - police killed the gunman - police kill the gunman
ROUGE-L Basically, an “F-measure” (or combination) of two normalized LCSs when 1.Again, no consecutive matches necessary 2.automatically includes longest in-sequence common n-gram.
ROUGE-L Reference two candidates ROUGE-L = 3/5 ROUGE-L = 1/2
ROUGE-L Basically, an “F-measure” (or combination) of two normalized LCSs when 1.Again, no consecutive matches necessary 2.automatically includes longest in-sequence common n-gram. 3.problem: counts only main in-sequence words, other LCSs and shorter CSs are not counted
Computing: ROUGE-W score so that consecutive matches should be awarded more than non-consecutive matches.
ROUGE-S Another “F-measure” but here using skip-bigram co-occurance statistics (i.e., non-consecutive but same order bi-grams). Goal is to measure overlap of skip-bigrams. We use function SKIP2(X,Y) to measure number of common skip-bigrams in X and Y.
ROUGE-S Using the SKIP2() function: No consecutive matches required, but still respects word order counts *all* in-order matching word pairs (LCS only counts longest common subsequence) Can impose limit on max skip distance –ROUGE-Sn, has max skip distance of n (e.g., ROUGE-S4)
Setup ISI’s A1Temp System 2002 NIST Chinese-English evaluation corpus 872 source sentences, 4 ref trans. each 1024-best lists used
Evaluating BLEU with ORANGE smoothed BLEU:
Evaluating BLEU with CC
Evaluating ROUGE-L/W with
Evaluating ROUGE-S with
Summary of metrics with