MultEval takes machine translation hypotheses from several runs of an optimizer and provides 3 popular metric scores, as well as, standard deviations (via bootstrap resampling) and p-values (via approximate randomization). This allows researchers to mitigate some of the risk of using unstable optimizers such as MERT, MIRA, and MCMC. It is intended to help in evaluating the impact of in-house experimental variations on translation quality; it is currently not setup to do bake-off style comparisons (bake-offs can't require multiple optimizer runs nor a standard tokenization).
It is a user-friendly implementation of: Jonathan Clark, Chris Dyer, Alon Lavie, and Noah Smith, "Better Hypothesis Testing for Statistical Machine Translation: Controlling for Optimizer Instability", Proceedings of the Association for Computational Lingustics, 2011. PDF
To keep updated on new versions of this software, subscribe to our low-traffic announcement mailing list: http://groups.google.com/group/multeval-announce. All active users are encourated to subscribe.
First, download and unpack the program:
wget http://www.cs.cmu.edu/~jhclark/downloads/multeval-0.5.1.tgz tar -xvzf multeval-0.5.1.tgz
To evaluate a single system from the example data and get its BLEU, METEOR, and TER scores along with its standard deviation use:
./multeval.sh eval --refs example/refs.test2010.lc.tok.en.* \ --hyps-baseline example/hyps.lc.tok.en.baseline.opt* \ --meteor.language en
The first time you run this command, METEOR (and its sizable paraphrase tables) will be downloaded. Also, to help the user determine if any tokenization mismatch happened, MultEval also prints the top OOVs according to METEOR.
To compare several systems from the example data and get its BLEU, METEOR, and TER scores along with their standard deviations and p-values, use:
./multeval.sh eval --refs example/refs.test2010.lc.tok.en.* \ --hyps-baseline example/hyps.lc.tok.en.baseline.opt* \ --hyps-sys1 example/hyps.lc.tok.en.sys1.opt* \ --hyps-sys2 example/hyps.lc.tok.en.sys2.opt* \ --meteor.language en
If you'd also like 1) a Latex table at you can copy-paste into your paper and 2) the hypotheses from the median optimization run ranked by improvement/decline over your baseline system and 3) A list of sentence-level metric scores including submetrics such as BLEU precision and brevity, then run it like this:
./multeval.sh eval --refs example/refs.test2010.lc.tok.en.* \ --hyps-baseline example/hyps.lc.tok.en.baseline.opt* \ --hyps-sys1 example/hyps.lc.tok.en.sys1.opt* \ --hyps-sys2 example/hyps.lc.tok.en.sys2.opt* \ --meteor.language en \ --latex table.tex \ --rankDir rank \ --sentLevelDir sentLevel
All files should contain tokenized, lowercased, space-delimited sentences in UTF-8 encoding, one sentence per line. Unlike many metric implementations, MultEval does no tokenization or segmentation for you (see discussion below).
Generally, you should evaluate full forms (i.e. without word segmentation). For languages without a canonical notion of words (e.g. Chinese, Japanese), we recommend splitting all non-Latin characters (e.g. each character that is not part of a borrowed Western word, URL, etc. should be evaluated as its own word.)
For a more detailed description of the various METEOR options, please see http://github.com/mjdenkowski/meteor.
METEOR and its paraphrase tables will automatically be downloaded from the web the first time you run multeval.sh. They are not included in the initial download due to the large size (~200MB) of the paraphrase tables.
The ASCII table produced by multeval looks something like this:
n=3 BLEU (s_sel/s_opt/p) METEOR (s_sel/s_opt/p) TER (s_sel/s_opt/p) Length (s_sel/s_opt/p) baseline 18.5 (0.3/0.1/-) 29.3 (0.1/0.0/-) 65.7 (0.4/0.2/-) 107.5 (0.4/0.1/-) system 1 18.8 (0.3/0.3/0.00) 30.3 (0.1/0.1/0.00) 64.8 (0.4/0.6/0.00) 107.7 (0.3/0.7/0.09) system 2 18.5 (0.3/0.1/0.00) 29.3 (0.1/0.0/0.00) 65.7 (0.4/0.2/0.00) 107.5 (0.4/0.1/0.00)
A quick explanation of these numbers (see paper for details):
- s_sel: The variance due to test set SELection. This is calculated using bootstrap resampling for each optimizer run and this number reports the average variance over all optimizer runs.
- s_opt: The variance due to OPTimizer instability. This is calculated directly as the variance of the aggregate metric score over all optimizer runs.
- p: This is the p-value calculated by approximate randomization. It can roughly be interpreted as the probability of the absolute difference between the baseline system and system i occurring due to chance where random permutations between the two systems are used to simulate chance occurrences. The quality of this measure depends on the n separate optimization runs of your system and is conditioned on your test set. See below for a more in-depth discussion on p-values.
The LaTeX table produced by multeval looks something like this:
To see a full list of options, use:
Usage: program <module_name> [options...] === TER === -T [--ter.shiftCost] Shift cost for TER -d [--ter.maxShiftDistance] Maximum shift distance for TER -P [--ter.punctuation] Use punctuation in TER? -b [--ter.beamWidth] Beam width for TER -B [--ter.substituteCost] Substitute cost for TER -D [--ter.deleteCost] Delete cost for TER -M [--ter.matchCost] Match cost for TER -I [--ter.insertCost] Insert cost for TER === BLEU === === METEOR === -t [--meteor.task] One of: rank adq hter tune (Rank is generally a good choice) -s [--meteor.synonymDirectory] If default is not desired (NOTE: This option has a different short flag than stand-alone METEOR) [optional] -x [--meteor.beamSize] Specify beam size (overrides default) -p [--meteor.params] Custom parameters of the form 'alpha beta gamma' (overrides default) [optional] -w [--meteor.weights] Specify module weights (overrides default) [optional] -a [--meteor.paraphraseFile] If default is not desired [optional] -m [--meteor.modules] Specify modules. (overrides default) Any of: exact stem synonym paraphrase [optional] -k [--meteor.keepPunctuation] Consider punctuation when aligning sentences (if false, the meteor tokenizer will be run, after which punctuation will be removed) -l [--meteor.language] Two-letter language code of a supported METEOR language (e.g. 'en') === MultEvalModule (for eval module) === -b [--boot-samples] Number of bootstrap replicas to draw during bootstrap resampling to estimate standard deviation for each system -H [--hyps-sys] Space-delimited list of files containing tokenized, fullform hypotheses, one per line -s [--ar-shuffles] Number of shuffles to perform to estimate p-value during approximate randomization test system *PAIR* -r [--rankDir] Rank hypotheses of median optimization run of each system with regard to improvement/decline over median baseline system and output to the specified directory for analysis [optional] -R [--refs] Space-delimited list of files containing tokenized, fullform references, one per line -o [--metrics] Space-delimited list of metrics to use. Any of: bleu, meteor, ter, length -F [--fullLatexDoc] Output a fully compilable Latex document instead of just the table alone [optional] -L [--latex] Latex-formatted table including measures that are commonly (or should be commonly) reported [optional] -D [--debug] Show debugging output? [optional] -B [--hyps-baseline] Space-delimited list of files containing tokenized, fullform hypotheses, one per line -v [--verbosity] Verbosity level --help help message
What do p-values actually mean?
A p-value is a model's estimate (where the model is a significance test) that a particular difference in scores arose by chance. Multeval uses approximate randomization, a test that approximates a permutation test via sampling shufflings of like hypotheses between systems.
The most important points are:
- a p-value does tell you whether a difference of this magnitude is likely to be generated again by some random process (a randomized optimizer)
- a p-value does not tell you whether a difference of this magnitude is meaningful (in terms of translation quality)
So even though a large difference may more frequently correspond to smaller p-values, this is not guaranteed. In fact, small differences can be quite significant and vice versa. For example, if you give a single optimizer sample with identical hypotheses and tell MultEval that these are actually two different systems (as in the baseline system and system 2 in the example data), there will be zero difference in scores and also a p-value of zero, since shuffling hypotheses between the systems produces no change, indicating that this difference (of zero) is likely to be reproducible. This demonstrates 2 points about p-values: 1) that this significance test does not account for the user giving it too few (optimizer) samples, which is why it's important to report how many optimizer samples you used and 2) that the test only provides information about the replicability of a delta, not whether or not the magnitude can be assigned external meaning (in terms of translation quality).
And here's a few FAQ's that come up from this discussion:
Why is the p-value ~= 0 for equal inputs
Because the quantity being estimated is actually p_ByChance(Diff=0). Given a single run of the optimizer, if two experimental configurations produce exactly the same output, we should (under frequentist statistics) say that this did not happen by chance. But of course we should distrust experiments with 1 optimizer run per experimental configuration. Now let's say you run each configuration 3 times and still both systems produce exactly the same output. Well, that's a very stable optimizer and we should begin to believe that this difference of zero has a very low probability of occurring by chance.
Why is the p-value ~= 0 for very different inputs?
The hypotheses before and after shuffling will most likely remain very different (as in the extreme case of comparing system1=MT, system2=references), so the probability of this difference occurring due to chance is quite low.
Why does the number of shuffles seem to affect p-value so much?
A small number of shuffles is just giving you a poor estimate with high variance and it's not advisable to run a small number of shuffles. Going above the default of 10k is fine, but it shouldn't change the p-values too much. Increasing the number of shuffles is actually just bringing you closer to the "true"estimate of zero, which we should expect from above.
So when should I expect non-zero p-values?
When shuffling produces strictly greater absolute differences than observed in aggregate and when it does this a non-trivial number of times. There's a few examples above illustrating this.
So I should accept papers with zero p-values?
Often not. This p-value is intended to estimate whether the difference is repeatable. But the magnitude of the difference still may not be substantive (users may not perceive any quality improvement). Sometimes, it may be more informative to look at the two sigma numbers calculated by multeval (test set variance and optimizer variance). Desite p-values and significance testing being a strong tradition in the statistics community, it does not provide a complete picture -- you need more than just the p-value. For example, it's very important to know the number of optimizer replicas.
MultEval generally rounds to one decimal place for metrics and two decimal places for p-values just before externally reporting values in tables. One should expect that our current evaluation measures and significance testing models are not likely to be very discriminant beyond this point and so assigning meaning to finer distinctions is not necessarily informative.
Discussion on Pitfalls in Tokenization, Segmentation, etc.
While tokenization can introduce bias in many metrics by changing the length of hypotheses and references, noise can also be introduced into the evaluation process by strange interactions between running a detokenizer followed by tokenizer. Therefore, since research often isn't focused on details such as tokenization, we recommend evaluating on full-form references and hypotheses (i.e. not segmented, unless you're working with a language without a canonical notion of words) that are tokenized. Of course, in a bake-off scenario or in research in which multiple tokenization schemes are being compared, a standard tokenization will be necessary to have a comparable evaluation.
Still, special care must be taken for issues such as segmentation, that can effect matching of resources such as paraphrase tables in METEOR. Since METEOR doesn't use paraphrases including punctuation, this isn't an issue for tokenization, while lowercased full-forms remain a requirement for METEOR evaluation.
Care must also be taken to ensure experimental validity so that a particular tokenization does not bias your results in an unexpected way. For example, it would be easy to inflate the BLEU score by segmenting URLs into many tokens (since URLs are usually passed through, you would almost always get credit for lots of extra tokens being correct). Now, your absolute scores will be higher and any changes (positive or negative) will be diluted since there is now a greater number of overall tokens, this uninteresting subset now makes up a greater overall fraction of the evaluation set.
As a final pitfall example, consider the evaluation of Buckwalter-transliterated Arabic data. Using the traditional NIST mt-eval script where tokenization is forced (and the final tokenization is never shown to the user), one would see very biased (nearly senseless) scores since the many punctuation characters used to encode the Arabic alphabet will be used to split should-be single tokens into many illogical pieces. In this case, traditional tokenization is almost certainly the wrong thing to do. This is why MultEval leaves tokenization to the user.
Cased vs Uncased Evaluation
For now, this program assumes you will give it lowercased input. Cased variants of BLEU and TER will be coming soon.
Using MultEval for Error Analysis and Oracle Scoring
In addition to metrics and statistical analyses for quantitative analysis, MultEval comes with several tools to help you qualitatively determine how your translation systems are doing. First, is the ability to rank hypotheses with regard to its improvement or decline over the baseline system (command line given above). Note, some metrics (e.g. BLEU) are notoriously unstable at the sentence level.
Second, is the ability to take a n-best list from a decoder (e.g. Moses, cdec), score all of the hypotheses using all of the metrics in MultEval, and then sort the hypotheses for each sentence by each of the metrics so that the first sentence output for each sentence is the n-best oracle. You can get this by running:
./multeval.sh nbest --nbest example/cdec.kbest \ --refs example/cdec.ref* \ --meteor.language en \ --rankDir rank \ > kbest.scored
MultEval will also display the corpus-level oracle score over the n-best list according to each metric.
For even more detailed analysis, you should also consider using the METEOR X-Ray analysis tool.
Compatibility and Accuracy
MultEval produces exactly the same metric scores you would obtain by running the metrics as stand-alone programs -- with the exception that MultEval does not perform tokenization. MultEval calls METEOR and TER using library API calls, and we regression test against gold-standard scores produced by these metrics in stand-alone mode. Its internal jBLEU library is a direct port of NIST's mteval-v13a.pl; we regression test against mteval-v13m.pl (which allows disabling normalization), included in the METEOR distribution. jBLEU implements the segment-level smoothing of BLEU from mteval-v13a.pl, which is derived from the smoothing in Kishore Papineni's bleu-1.04.pl script (the original IBM implementation of BLEU), added as of 3/9/2004. Note this is not the simplistic +1 BLEU smoothing used in some other system. Instead, bleu-1.04.pl uses an exponential decay function to give non-zero credit for orders that have no matches; this makes a difference only when some order has zero matches, which is relatively rare for corpus-level BLEU, but may matter for sentence-level BLEU.
You can see this reflected in most of MultEval's regression tests, which come out the same for both Moses' multi-bleu.pl and mteval-v13a.pl: http://travis-ci.org/jhclark/multeval
Comparison with Moses' multi-bleu.pl
Moses multi-bleu.pl calculates BLEU in a slightly different way than MultEval: MultEval uses the "smoothed" variant of BLEU in which orders with zero matching n-grams but non-zero possible matches get smoothed as per a formula in Papineni's bleu-1.04.pl.
Adding Your Own Metric
Implement the metrics.Metric interface and then add it as an option in multeval.MultEval.
MultEval uses the following libraries:
- METEOR 1.4 (LGPL License, http://www.cs.cmu.edu/~alavie/METEOR/ -- WordNet database has a compatible free license)
- Translation Error Rate 0.7 (LGPL License, TerCom, http://www.cs.umd.edu/~snover/tercom/ -- Thanks to Matt Snover for relicensing this). MultEval uses a slightly modified version of TER; the only change is that all classes have been moved from the default package into a "ter" package to overcome the Java restriction that classes in the defalut package cannot be imported across JARs.
- Google Guava (Apache License)
- Java Annotation Options (jannopts, LGPL License)
Should you want to build MultEval yourself instead of using the provided tarball distribution, you'll need to download meteor using get_deps.sh. Then you can just run ant:
$ ./get_deps.sh # Download meteor $ ant
NOTE: There's a strange generics-related javac bug that's known to cause the build to fail under OpenJDK V1.6.0_17. However, this seems to be resolved as of version 1.6.0_21.
If you use this software, consider citing:
Jonathan Clark, Chris Dyer, Alon Lavie, and Noah Smith, "Better Hypothesis Testing for Statistical Machine Translation: Controlling for Optimizer Instability", Proceedings of the Association for Computational Lingustics, 2011.
The included metrics should be cited as:
Kishore Papineni, Salim Roukos, Todd Ard, and Wei-Jing Zhu, "BLEU: a method for automatic evaluation of machine translation," Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, 2002.
Michael Denkowski and Alon Lavie, "Meteor 1.3: Automatic Metric for Reliable Optimization and Evaluation of Machine Translation Systems," Proceedings of the EMNLP 2011 Workshop on Statistical Machine Translation, 2011. (NOTE: MultEval uses Meteor 1.4; Meteor 1.4 is the same as Meteor 1.3, but just has support for more languages)
Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul, "A Study of Translation Edit Rate with Targeted Human Annotation," Proceedings of Association for Machine Translation in the Americas, 2006.