geval/README.md

GEval

GEval is a Haskell library and a stand-alone tool for evaluating the results of solutions to machine learning challenges as defined on the Gonito platform. Also could be used outside the context of Gonito.net challenges, assuming the test data is given in simple TSV (tab-separated values) files.

Note that GEval is only about machine learning evaluation. No actual machine learning algorithms are available here.

The official repository is git://gonito.net/geval, browsable at https://gonito.net/gitlist/geval.git/.

Installing

You need Haskell Stack. You could install Stack with your package manager or with:

curl -sSL https://get.haskellstack.org/ | sh

When you've got Haskell Stack, install GEval with:

git clone git://gonito.net/geval
cd geval
stack setup
stack test
stack install

(Note that when you're running Haskell Stack for the first time it will take some time and a couple of gigabytes on your disk.)

By default, geval binary is installed in $HOME/.local/bin, so in order to run geval you need to either add $HOME/.local/bin to $PATH in your configuration or to type:

PATH="$HOME/.local/bin" geval ...

Plan B — just download the GEval binary

(Assuming you have a standard 64-bit Linux.)

wget https://gonito.net/get/bin/geval
chmod u+x geval
./geval --help

Quick tour

Let's use GEval to evaluate machine translation (MT) systems (but keep in mind than GEval could be used for many other machine learning task types). We start with simple evaluation, but then we switch to what might be called black-box debugging of ML models.

First, we will run GEval on WMT-2017, a German-to-English machine translation challenge repackaged for Gonito.net platform and available there (though, in a moment you'll see it can be run on other test sets, not just the ones conforming to specific Gonito.net standards). Let's download one of the solutions, it's just available via git, so you don't have to click anywhere, just type:

git clone git://gonito.net/wmt-2017 -b submission-01229

Let's step into the repo and run GEval (I assume you added geval path to $PATH, so that you could just use geval instead of /full/path/to/geval):

cd submission-01229
geval

Well, something apparently went wrong:

geval: No file with the expected results: `./test-A/expected.tsv`

The problem is that the official test set is hidden from you (although you can find it if you are determined...) You should try running GEval on the dev set instead:

geval -t dev-0

and you'll see the result — 0.27358 in BLEU metric, which is the default metric for the WMT-2017 challenge. GEval could do the evaluation using other metrics, in case of machine translation, (Google) GLEU (alternative to BLEU), WER (word-error rate) or simple accuracy (which could be interpreted as sentence-recognition rate here) might make sense:

geval -t dev-0 --metric GLEU --metric WER --metric Accuracy

After a moment, you'll see the results:

BLEU	0.27358
GLEU	0.31404
WER	0.55201
Accuracy	0.01660

Ah, we forgot about the tokenization, in order to properly calculate BLEU (or GLEU) the way it was done within the official WMT-2017 challenge, you need to tokenize the expected output and the actual output of your system using the right tokenizer:

geval -t dev-0 --metric GLEU --metric WER --metric Accuracy --tokenizer 13a

BLEU	0.26901
WER 0.58858
GLEU	0.30514
Accuracy	0.01660

The results do not look good anyway and I'm not talking about Accuracy, which, even for a good MT (or even a human), will be low (as it measures how many translations are exactly the same as the golden standard), but rather about BLEU which is not impressive for this particular task. Actually, it's no wonder as the system we're evaluating now is a very simple neural machine translation baseline. Out of curiosity, let's have a look at the worst items, i.e. sentences for which the GLEU metric is the lowest (GLEU is better than BLEU for item-per-item evaluation); it's easy with GEval:

geval -t dev-0 --alt-metric GLEU --line-by-line --sort | head -n 10

0.0	Tanzfreudiger Nachwuchs gesucht	Dance-crazy youths wanted	Dance joyous offspring sought
0.0	Bulgarische Gefängnisaufseher protestieren landesweit	Bulgaria 's Prison Officers Stage National Protest	Bulgarian prison guards protest nationwide
0.0	Schiffe der Küstenwache versenkt	Coastguard ships sunk	Coast Guard vessels sinking
0.0	Gebraucht kaufen	Buying used	Needed buy
0.0	Mieten	Renting	Rentals
0.0	E-Books	E-books	E-Books
0.021739130434782608	Auch Reservierungen in Hotels gehen deutlich zurück.	There is even a marked decline in the number of hotel reservations .	Reservations also go back to hotels significantly .
0.023809523809523808	Steuerbelastung von Geschäftsleuten im Raum Washington steigt mit der wirtschaftlichen Erholung	Washington-area business owners " tax burden mounts as economy rebounds	Tax burden of businessmen in the Washington area rises with economic recovery
0.03333333333333333	Verwunderte Ärzte machten Röntgenaufnahmen seiner Brust und setzen Pleurakathether an, um Flüssigkeit aus den Lungen zu entnehmen und im Labor zu testen.	Puzzled doctors gave him chest X-rays , and administered pleural catheters to draw off fluid from the lungs and send it for assessment .	At the end of his life , she studied medicine at the time .
0.03333333333333333	Die Tradition der Schulabschlussbälle in den USA wird nun auf die Universitäten übertragen, wo Freshmen Auftritte mit dem Privatflugzeug angeboten werden.	US prom culture hits university life with freshers offered private jet entrances	The tradition of school leavers in the U.S. is now transferred to universities , where freshmen are offered appearances with the private plane .

Well, this way, we found some funny utterances for which even a single word was recovered, but could we get more insight?

The good news is that you could use GEval to debug the MT system in a black-box manner to order to find its weak points -- --worst-features is the option to do this:

geval -t dev-0 --alt-metric GLEU --worst-features | head -n 10

This command will find the top 10 "worst" features (in either input, expected output or actual output), i.e. the features which correlate with low GLEU values in the most significant way.

exp:"	346	0.27823151	0.00000909178949766883
out:''	348	0.28014113	0.00002265047322460752
exp:castle	23	0.20197660	0.00006393156973075869
exp:be	191	0.27880383	0.00016009575605100586
exp:road	9	0.16307514	0.00025767878872874620
exp:out	78	0.26033671	0.00031551452260174863
exp:(	52	0.25348798	0.00068739029500072100
exp:)	52	0.25386216	0.00071404713888387060
exp:club	28	0.22958093	0.00078051481428704770
out:`	9	0.17131601	0.00079873676961809170

How to read the output like this?

1. The feature (i.e. a word or token) found, prepended with a qualifier: exp for the expected output, out — the actul output, in — input.
2. Number of occurrences.
3. The mean score for all items (in our examples: sentences) with a given feature. For instance, the average GLEU score for sentences for which a double quote is expected is 0.27823151. At first glance, it does not seem much worse than the general score (0.30514), but actually…
4. … it's highly significant. The probability to get it by chance (according to Mann-Whitney U test) is extremely low (p = 0.000009).

But why were double quotes so problematic in German-English translation?! Well, look at the second worst feature — '' in the output! Oops, it seems like a very stupid mistake with post-processing was done and no double quote was correctly generated, which decreased the score a little bit for each sentence in which the quote was expected.

When I fixed this simple bug, the BLUE metric increased from 0.27358 to 0.27932!

What about the third item — the word castle in the expected output? Let's have a look at the examples with this word using --line-by-line option combined with grep:

geval -t dev-0 --alt-metric GLEU --line-by-line --sort | grep 'castle' | head -n 5

0.0660377358490566	Eine Wasserburg, die bei unserer nächsten Aufgabe gesucht wird, ist allerdings in der Höhe eher selten zu finden.	A moated castle , which we looked for as part of our next challenge , is , of course , rather hard to find way up high .	However , a watershed that is being sought in our next assignment is rather rare .
0.07142857142857142	Ziehen die Burgvereine bald wieder an einem Strang?	Will the Burgvereine ( castle clubs ) get back together again ?	Do the Burgundy clubs join forces soon ?
0.11290322580645161	Zuletzt gab es immer wieder Zwist zwischen den beiden Wolfratshauser Burgvereinen.	Recently there have been a lot of disputes between both of the castle groups in Wolfratshausen .	Last but not least , there has been a B.A. between the two Wolfratshauser Burgundy .
0.11650485436893204	Während die Burgfreunde um den plötzlich verstorbenen Richard Dimbath bis zuletzt einen Wiederaufbau der Burg am Bergwald im Auge hatten, steht für den Burgverein um Sjöberg die "Erschließung und Erlebbarmachung" des Geländes an vorderster Stelle.	Whereas the castle friends , and the recently deceased Richard Dimbath right up until the bitter end , had their eyes on reconstructing the castle in the mountain forest , the castle club , with Sjöberg , want to " develop and bring the premises to life " in its original place .	While the castle fans were aware of the sudden death of Richard Dimbath until the end of a reconstruction of the castle at Bergwald , the Burgverein around Sjöberg is in the vanguard of the `` development and adventure '' of the area .
0.1206896551724138	Auf der Hüpfburg beim Burggartenfest war am Sonnabend einiges los.	Something is happening on the bouncy castle at the Burggartenfest ( castle garden festival ) .On the edge of the castle there was a lot left at the castle castle .

Well, now it is not as simple as the problem with double quotes. It seems that "castle" German is full of compounds which are hard for the MT system analysed, in particular the word Burgverein makes the system trip up. You might try to generalise this insight and improve your system or you might not. It might be considered an issue in the test set rather than in the system being evaluated. (Is it OK that we have so many sentences with Burgverein in the test set?)

But do you need to represent your test set a Gonito challenge to run GEval? Actually no, I'll show this by running GEval directly on WMT-2018. First, let's download the files:

wget http://data.statmt.org/wmt17/translation-task/wmt17-submitted-data-v1.0.tgz
tar vxf wmt17-submitted-data-v1.0.tgz

and run GEval for one of the submissions (UEdin-NMT):

geval --metric BLEU --precision 4 --tokenizer 13a \
    -i wmt17-submitted-data/txt/sources/newstest2017-deen-src.de \
    -o wmt17-submitted-data/txt/system-outputs/newstest2017/de-en/newstest2017.uedin-nmt.4723.de-en \
    -e wmt17-submitted-data/txt/references/newstest2017-deen-ref.en

0.3430

where -i stands for the input file, -o — output file, -e — file with expected (reference) data.

Let's evaluate another system:

geval --metric BLEU --precision 4 --tokenizer 13a \
    -i wmt17-submitted-data/txt/sources/newstest2017-deen-src.de \
    -o wmt17-submitted-data/txt/system-outputs/newstest2017/de-en/newstest2017.LIUM-NMT.4733.de-en \
    -e wmt17-submitted-data/txt/references/newstest2017-deen-ref.en

0.2939

In general, LIUM is much worse than UEdin, but were there any utterance for which UEdin is worse than LIUM? You could use --diff option to find this:

geval --metric GLEU --precision 4 --tokenizer 13a \
    -i wmt17-submitted-data/txt/sources/newstest2017-deen-src.de \
    -o wmt17-submitted-data/txt/system-outputs/newstest2017/de-en/newstest2017.uedin-nmt.4723.de-en \
    --diff wmt17-submitted-data/txt/system-outputs/newstest2017/de-en/newstest2017.LIUM-NMT.4733.de-en \
    -e wmt17-submitted-data/txt/references/newstest2017-deen-ref.en -s | head -n 10

The above command will print out the 10 sentences for which the difference between UEdin and LIUM is the largest:

-0.5714285714285714     Hier eine Übersicht:    Here is an overview:    Here is an overview:    Here's an overview:
-0.5714285714285714     Eine Generation protestiert.    A generation is protesting.     A generation is protesting.     A generation protesting.
-0.5102564102564102     Bald stehen neue Container in der Wasenstraße   New containers will soon be located in Wasenstraße      New containers will soon be available on Wasenstraße    Soon, new containers are in the water road
-0.5    "Die ersten 100.000 Euro sind frei."    "The first 100,000 euros are free."     "The first 100.000 euros are free."     'the first £100,000 is free. '
-0.4736842105263158     Als gefährdet gelten auch Arizona und Georgia.  Arizona and Georgia are also at risk.   Arizona and Georgia are also at risk.   Arizona and Georgia are also considered to be at risk.
-0.4444444444444445     Das ist alles andere als erholsam.      This is anything but relaxing.  That is anything but relaxing.  This is far from relaxing.
-0.4285714285714286     Ein Haus bietet Zuflucht.       One house offers refuge.        A house offers refuge.  A house offers sanctuary.
-0.42307692307692313    Weshalb wir Simone, Gabby und Laurie brauchen   Why we need Simone, Gabby and Laurie    Why we need Simone, Gabby and Laurie    Why We Need Simone, Gabby and Laurie
-0.4009009009009009     Die "Identitäre Bewegung" ist eine Gruppierung mit französischen Wurzeln, die seit 2012 auch in Deutschland aktiv ist.  The "Identitäre Bewegung" is a group with French roots that has been active in Germany since 2012.     The "identitarian movement" is a group with French roots that has been active in Germany since 2012.    The "Identitarian Movement" is a grouping with French roots, which has also been active in Germany since 2012.
-0.4004524886877827     Der Mann soll nicht direkt angesprochen werden. The man should not be approached.       The man should not be addressed directly.       The man is not expected to be addressed directly.

The columns goes as follows:

1. the difference between the two systems (GLEU "delta")
2. input
3. expected output (reference translation)
4. the output from LIUM
5. the output from UEdint

Hmmm, turning 100.000 euros into £100,000 is no good…

You could even get the list of the "most worsening" features between LIUM and UEdin, the features which were "hard" for UEdin, even though they were easy for UEdin:

geval --metric GLEU --precision 4 --tokenizer 13a \
  -i wmt17-submitted-data/txt/sources/newstest2017-deen-src.de \
  -o wmt17-submitted-data/txt/system-outputs/newstest2017/de-en/newstest2017.uedin-nmt.4723.de-en \
  --most-worsening-features wmt17-submitted-data/txt/system-outputs/newstest2017/de-en/newstest2017.LIUM-NMT.4733.de-en \
  -e wmt17-submitted-data/txt/references/newstest2017-deen-ref.en | head -n 10

exp:euros       31      -0.06888191     0.00000597009369023376
in<1>:Euro      31      -0.05745546     0.00001601359624974303
exp:be  295     0.01980564      0.00039919436507962510
exp:Federal     12      -0.05519148     0.00044208689767323860
exp:small       21      -0.02782365     0.00097178111718141370
exp:40  9       -0.05635973     0.00121990447157221800
in<1>:40        9       -0.05635973     0.00121990447157221800
out:interior    6       -0.07150121     0.00132944903870436640
exp:turnover    9       -0.09077533     0.00147928107739624940
exp:head        17      -0.03198173     0.00170431081987969600

Hey, UEdin, you have a problem with euros… is it due to Brexit?

Another example

Let us download a Gonito.net challenge:

git clone git://gonito.net/sentiment-by-emoticons

The task is to predict the sentiment of a Polish short text -- whether it is positive or negative (or to be precise: to guess whether a positive or negative emoticon was used). The train set is given in the train/train.tsv.xz file, each item is given in a separate file, have a look at the first 5 items:

xzcat train/train.tsv.xz | head -n 5

Now let's try to evaluate some solution to this challenge. Let's fetch it:

git fetch git://gonito.net/sentiment-by-emoticons submission-01865
git reset --hard FETCH_HEAD

and now run geval:

geval -t dev-0

(You need to run dev-0 test as the expected results for the test-A test is hidden from you.) The evaluation result is 0.47481. This might be hard to interpret, so you could try other metrics.

geval -t dev-0 --metric Accuracy --metric Likelihood

So now you can see that the accuracy is over 78% and the likelihood (i.e. geometric mean of probabilities of the correct classes) is 0.62.

Preparing a Gonito challenge

Directory structure of a Gonito challenge

A definition of a Gonito challenge should be put in a separate directory. Such a directory should have the following structure:

• README.md — description of a challenge in Markdown, the first header will be used as the challenge title, the first paragraph — as its short description
• config.txt — simple configuration file with options the same as the ones accepted by geval binary (see below), usually just a metric is specified here (e.g. --metric BLEU), also non-default file names could be given here (e.g. --test-name test-B for a non-standard test subdirectory)
• train/ — subdirectory with training data (if training data are supplied for a given Gonito challenge at all)
• train/train.tsv — the usual name of the training data file (this name is not required and could be more than one file), the first column is the target (predicted) value, the other columns represent features, no header is assumed
• dev-0/ — subdirectory with a development set (a sample test set, which won't be used for the final evaluation)
• dev-0/in.tsv — input data (the same format as train/train.tsv, but without the first column)
• dev-0/expected.tsv — values to be guessed (note that paste dev-0/expected.tsv dev-0/in.tsv should give the same format as train/train.tsv)
• dev-1/, dev-2, ... — other dev sets (if supplied)
• test-A/ — subdirectory with the test set
• test-A/in.tsv — test input (the same format as dev-0/in.tsv)
• test-A/expected.tsv — values to be guessed (the same format as dev-0/expected.tsv), note that this file should be “hidden” by the organisers of a Gonito challenge, see notes on the structure of commits below
• test-B, test-C, ... — other alternative test sets (if supplied)

Initiating a Gonito challenge with geval

You can use geval to initiate a Gonito challenge:

geval --init --expected-directory my-challenge

(This will generate a sample toy challenge about guessing planet masses).

A metric (other than the default RMSE — root-mean-square error) can be given to generate another type of toy challenge:

geval --init --expected-directory my-machine-translation-challenge --metric BLEU

Preparing a Git repository

Gonito platform expects a Git repository with a challenge to be submitted. The suggested way to do this is as follows:

1. Prepare a branch with all the files without test-A/expected.tsv. This branch will be cloned by people taking up the challenge.
2. Prepare a separate branch (or even a repo) with test-A/expected.tsv added. This branch should be accessible by Gonito platform, but should be kept “hidden” for regular users (or at least they should be kindly asked not to peek there). It is recommended (though not obligatory) that this branch contain all the source codes and data used to generate the train/dev/test sets. (Use git-annex if you have really big files there.)

Branch (1) should be the parent of the branch (2), for instance, the repo (for the toy “planets” challenge) could be created as follows:

geval --init --expected-directory planets
cd planets
git init
git add .gitignore config.txt README.md train/train.tsv dev-0/{in,expected}.tsv test-A/in.tsv
git commit -m 'init challenge'
git remote add origin ssh://gitolite@gonito.net/filipg/planets
git push origin master
git branch dont-peek
git checkout dont-peek
git add test-A/expected.tsv
git commit -m 'with expected results'
git push origin dont-peek

Taking up a Gonito challenge

Clone the repo with a challenge, as given on the Gonito web-site, e.g. for the toy “planets” challenge (as generated with geval --init):

git clone git://gonito.net/planets

Now use the train data and whatever machine learning tools you like to guess the values for the dev set and the test set, put them, respectively, as:

• dev-0/out.tsv
• test-A/out.tsv

(These files must have exactly the same number of lines as, respectively, dev-0/in.tsv and test-0/in.tsv. They should contain only the predicted values.)

Check the result for the dev set with geval:

geval --test-name dev-0

(the current directory is assumed for --out-directory and --expected-directory).

If you'd like and if you have access to the test set results, you can “cheat” and check the results for the test set:

cd ..
git clone git://gonito.net/planets planets-secret --branch dont-peek
cd planets
geval --expected-directory ../planets-secret

Uploading your results to Gonito platform

Uploading is via Git — commit your “out” files and push the commit to your own repo. On Gonito you are encouraged to share your code, so be nice and commit also your source codes.

git remote add mine git@github.com/johnsmith/planets-johnsmith
git add {dev-0,test-A}/out.tsv
git add Makefile magic-bullet.py ... # whatever scripts/source codes you have
git commit -m 'my solution to the challenge'
git push mine master

Then let Gonito pull them and evaluate your results, either manually clicking "submit" at the Gonito web site or using --submit option (see below).

Submitting a solution to a Gonito platform with GEval

A solution to a machine learning challenge can be submitted with the special --submit option:

geval --submit --gonito-host HOST --token TOKEN

where:

• HOST is the name of the host with a Gonito platform
• TOKEN is a special per-user authorisation token (can be copied from "your account" page)

HOST must be given when --submit is used (unless the creator of the challenge put --gonito-host option in the config.txt file, note that in such a case using --gonito-host option will result in an error).

If TOKEN was not given, GEval attempts to read it from the .token file, and if the .token file does not exist, the user is asked to type it (and then the token is cached in .token file).

GEval with --submit does not commit or push changes, this needs to be done before running geval --submit. On the other hand, GEval will check whether the changes were committed and pushed.

Note that using --submit option for the main instance at https://gonito.net is usually NOT needed, as the git repositories are configured there in such a way that an evaluation is triggered with each push anyway.

geval options

geval - stand-alone evaluation tool for tests in Gonito platform

Usage: geval ([--init] | [-v|--version] | [-l|--line-by-line] |
             [-w|--worst-features] | [-d|--diff OTHER-OUT] |
             [-m|--most-worsening-features ARG] | [-j|--just-tokenize] |
             [-S|--submit]) ([-s|--sort] | [-r|--reverse-sort])
             [--out-directory OUT-DIRECTORY]
             [--expected-directory EXPECTED-DIRECTORY] [-t|--test-name NAME]
             [-o|--out-file OUT] [-e|--expected-file EXPECTED]
             [-i|--input-file INPUT] [-a|--alt-metric METRIC]
             [-m|--metric METRIC] [-p|--precision NUMBER-OF-FRACTIONAL-DIGITS]
             [-T|--tokenizer TOKENIZER] [--gonito-host GONITO_HOST]
             [--token TOKEN]
  Run evaluation for tests in Gonito platform

Available options:
  -h,--help                Show this help text
  --init                   Init a sample Gonito challenge rather than run an
                           evaluation
  -v,--version             Print GEval version
  -l,--line-by-line        Give scores for each line rather than the whole test
                           set
  -w,--worst-features      Print a ranking of worst features, i.e. features that
                           worsen the score significantly. Features are sorted
                           using p-value for Mann-Whitney U test comparing the
                           items with a given feature and without it. For each
                           feature the number of occurrences, average score and
                           p-value is given.
  -d,--diff OTHER-OUT      Compare results of evaluations (line by line) for two
                           outputs.
  -m,--most-worsening-features ARG
                           Print a ranking of the "most worsening" features,
                           i.e. features that worsen the score the most when
                           comparing outputs from two systems.
  -j,--just-tokenize       Just tokenise standard input and print out the tokens
                           (separated by spaces) on the standard output. rather
                           than do any evaluation. The --tokenizer option must
                           be given.
  -S,--submit              Submit current solution for evaluation to an external
                           Gonito instance specified with --gonito-host option.
                           Optionally, specify --token.
  -s,--sort                When in line-by-line or diff mode, sort the results
                           from the worst to the best
  -r,--reverse-sort        When in line-by-line or diff mode, sort the results
                           from the best to the worst
  --out-directory OUT-DIRECTORY
                           Directory with test results to be
                           evaluated (default: ".")
  --expected-directory EXPECTED-DIRECTORY
                           Directory with expected test results (the same as
                           OUT-DIRECTORY, if not given)
  -t,--test-name NAME      Test name (i.e. subdirectory with results or expected
                           results) (default: "test-A")
  -o,--out-file OUT        The name of the file to be
                           evaluated (default: "out.tsv")
  -e,--expected-file EXPECTED
                           The name of the file with expected
                           results (default: "expected.tsv")
  -i,--input-file INPUT    The name of the file with the input (applicable only
                           for some metrics) (default: "in.tsv")
  -a,--alt-metric METRIC   Alternative metric (overrides --metric option)
  -m,--metric METRIC       Metric to be used - RMSE, MSE, Accuracy, LogLoss,
                           Likelihood, F-measure (specify as F1, F2, F0.25,
                           etc.), multi-label F-measure (specify as
                           MultiLabel-F1, MultiLabel-F2, MultiLabel-F0.25,
                           etc.), MAP, BLEU, NMI, ClippEU, LogLossHashed,
                           LikelihoodHashed, BIO-F1, BIO-F1-Labels or CharMatch
  -p,--precision NUMBER-OF-FRACTIONAL-DIGITS
                           Arithmetic precision, i.e. the number of fractional
                           digits to be shown
  -T,--tokenizer TOKENIZER Tokenizer on expected and actual output before
                           running evaluation (makes sense mostly for metrics
                           such BLEU), minimalistic, 13a and v14 tokenizers are
                           implemented so far. Will be also used for tokenizing
                           text into features when in --worst-features and
                           --most-worsening-features modes.
  --gonito-host GONITO_HOST
                           Submit ONLY: Gonito instance location.
  --token TOKEN            Submit ONLY: Token for authorization with Gonito
                           instance.

If you need another metric, let me know, or do it yourself!

Licence

Apache License 2.0

Authors

Filip Graliński

Contributors

Piotr Halama