Addicter

Addicter stands for Automatic Detection and DIsplay of Common Translation ERrors. It will be a set of tools (mostly scripts written in Perl) that help with error analysis for machine translation.

The work on Addicter has started at the MT Marathon 2010 in Dublin, within a broader 5-day project called Failfinder (Dan Zeman, Ondřej Bojar, Martin Popel, David Mareček, Jon Clark, Ken Heafield, Qin Gao, Loïc Barrault). The code that resulted from the project can be freely downloaded from https://failfinder.googlecode.com/svn/trunk/. The nucleus that existed just after the MT Marathon (4 Feb 2010) is Addicter version 0.1, to reflect that this was by no means deemed a final product.

In 2011, the viewer was accompanied by an automatic error recognizer and classifier, thanks to Mark Fishel. The development has been moved to ÚFAL StatMT SVN repository (i.e. failfinder.googlecode.com is currently not maintained). In September 2011 at the Sixth MT Marathon in Trento, Addicter was further developed and thoroughly compared with another tool for error analysis, Hjerson. See the project wiki. For further developments, see also the Terra website.

Currently, Addicter can do the following:

• Find erroneous tokens and classify the errors in a way similar to Vilar's taxonomy.
• Browse the test data, sentence by sentence, and show aligned source sentence, reference translation and system hypothesis.
• Browse aligned training corpus, look for example words in context.
• Show lines of the phrase table that contain a given word.
• Summarize alignments of a given word. This feature can also serve as a primitive corpus-based dictionary.
• In the near future, we also plan to add searching and grouping of words sharing the same lemma. That way morphological errors will be highlighted.

The viewing and browsing is performed using a web server that generates web pages dynamically (to avoid pre-generating millions of static HTML documents). Words in sentences are clickable so that the user can quickly navigate to examples and summaries of other than the current word. If you have access to a webserver you may use Addicter with it; otherwise you can use Addicter's own lightweight server. A small subset can be also generated as static HTML files and viewed without a web server: the test data browser.

There is another subpage for Addicter in this wiki that lies in the external name space, thus it can be used for external collaboration.

• Addicter is written in Perl and you need a Perl interpreter to run Addicter. This is usually no problem on Unix-like systems but you may need to install Perl version ≥ 5.8 if you are working on Windows. Options include Active Perl and Strawberry Perl.
• There is now no need to install or have access to a web server; nevertheless, we keep the tutorial section below for those interested.
• To be able to generate alignments that will be displayed by Addicter, you need Giza++ or equivalent. The first training few steps of the Moses suite will do.
• Check out Addicter code from the ÚFAL SVN repository (see below how).

NOTE: Since September 2011, it is not necessary to install a local web server, so skip this section if you do not want it. Addicter now comes with a script called server.pl that works as a HTTP daemon and serves Addicter content (but nothing else) to your browser. This section is thus optional.

This tutorial currently focuses on installing Apache HTTP Server on Microsoft Windows. If you are experienced user of another operating system and wish to share advice, please feel free to contact me.

• Download the Apache HTTP Server from http://httpd.apache.org/download.cgi. For MS Windows, you can download a package for the Microsoft Installer (.msi). Install it by double-clicking on the installation file. I suggest installing Apache as a system service. That way, it will automatically start on startup of your computer.
• Configure the server. This essentially means editing a configuration file and restarting the server. Depending on your system settings, Apache version etc., the configuration file will reside in a path similar to this: C:\Program Files\Apache Software Foundation\Apache2.2\conf\httpd.conf. Alternatively, you can access it via your Start Menu: Apache → Apache HTTP Server 2.2 → Configure Apache Server → Edit the Apache httpd.conf Configuration File.
  • Look for a ScriptAlias directive. It tells the server: 1. what path on the hard disk contains scripts that can generate dynamic HTML content on the fly, and 2. how the path will be represented in the URL (web address). For example

    ScriptAlias /cgi/ "C:/Documents and Settings/Dan/Documents/Web/cgi/"

    says that the URL http://localhost/cgi/anyscript.pl leads to your script C:\Documents and Settings\Dan\Documents\Web\cgi\anyscript.pl, and that it's a script (i.e., the server shall invoke it and send its output, instead of sending the script itself).

  • Under Windows, you will also want to set

    ScriptInterpreterSource registry

    It tells the server that the Windows registry shall be used to figure out how to run a script (e.g., that C:\Perl\Perl.exe binary must be run to interpret a .pl script).

  • CGI scripts will not run under the same environment as a user command line. They will not see the PERLLIB variable and thus not find the libraries unless we specifically instruct Apache to pass the variable to the CGI environment:

    PassEnv PERLLIB PERL5LIB

• Restart the server. On the main Windows panel, there is (typically in the lower right corner) a set of icons, including a new one for Apache. Right-click on it, select Open Apache Monitor, then Restart.

Install the Apache HTTP server package. After successful installation, there should be a file /etc/apache2/sites-enabled/000-default. Edit it (you need root permissions). There should be a section similar to the following:

	ScriptAlias /cgi-bin/ /usr/lib/cgi-bin/
	<Directory "/usr/lib/cgi-bin">
		AllowOverride None
		Options +ExecCGI -MultiViews +SymLinksIfOwnerMatch
		Order allow,deny
		Allow from all
	</Directory>

Either create a copy of the section with new alias and path (eg. ScriptAlias /addicter-cgi/ /home/user/addicter/cgi/) or use the /usr/lib/cgi-bin (or whatever folder you see by default) for your addicter CGI scripts and data (see below).

We use $CGI to refer to the path you registered with Apache as containing CGI scripts (using the ScriptAlias directive). NOTE: If you are using Addicter's own web server or if Addicter content is the only thing you intend to use the server to serve, probably the easiest thing to do is to set the Addicter's cgi folder as your $CGI. NOTE 2: There are couple of files with static (non-CGI) web content, needed by the CGI scripts. These files (currently tabs.gif and activatables.js) are in $CGI/... With Addicter's own web server, this is just fine. If you are using another web server, however, you must copy these files to the appropriate location in your static content directory structure so that the server finds them. They should not be directly in the $CGI folder because they are not scripts and should not be treated as scripts by the server.

• Addicter uses some general-purpose Perl libraries that are maintained in a separate repository. Download these first, using username public and password public. Then make sure that Perl finds these libraries. In Linux/bash, the following commands will do that:

  svn --username public checkout https://svn.ms.mff.cuni.cz/svn/dzlib ~/lib
  export PERL5LIB=~/lib:$PERL5LIB

  In Windows, you can use TortoiseSVN to access the repository.

• Check out the current version of Addicter from the StatMT SVN repository, again using username public and password public:

  svn --username public checkout https://svn.ms.mff.cuni.cz/svn/statmt/trunk/addicter addicter

• There are three subfolders, testchamber, prepare and cgi. Copy the contents of the cgi folder to $CGI.
• For every experiment whose data shall be explored by addicter, create a subfolder in $CGI, e.g. $CGI/fr-en-exp01.

Before invoking the viewer, you need to run an indexing script over your aligned corpus. It will create a bunch of index files that will later tell the viewer where to look for examples of a particular word. The indexer needs the following input files:

• train.src … source side of training corpus
• train.tgt … target side of training corpus
• train.ali … alignment of training corpus
• test.src … source side of test data
• test.tgt … reference translation of test data
• test.ali … alignment of the source and reference translation of test data
• test.system.tgt … system output for test data
• test.system.ali … alignment of the source and the system output for test data

<!–The prepare folder contains some sample corpora in sample_data.zip.–>

The indexer splits the output index into multiple files in order to reduce size of any individual file. All index files must be stored in the experiment subfolder of $CGI so that the CGI scripts can find them.

We assume that your corpus is already sentence-aligned and tokenized. I.e., source and target files have the same number of lines (sentences, segments), and tokens (words, punctuation) on each line are space-separated. If you are using Addicter to perform analysis of errors made by a machine translation system, you probably already have such a corpus. You may also want to use a lowercased version of your corpus. Unless stated otherwise, all files are supposed to be plain text files in the UTF-8 encoding.

You will also need some alignment files that define bi-directional word alignments. If you have trained a statistical MT system such as Moses, chances are that you already have such files for the training data. They result from the first three steps of the Moses training pipeline, namely from two runs of Giza++ and an alignment symmetrization algorithm. In order to get alignments for test data, too, you can do the following:

• Join the source training file with the source test file. Similarly, join the target sides of the two data sets.
• Re-run Giza++ over the joint corpus.
• The resulting alignment file has the same number of lines as the source and the target side of the corpus. By cutting off the last N lines, you easily separate the training and test alignments from each other.
• Alternatively, you can use MGiza (presented at MT Marathon 2010 in Dublin) to align the test data. After the aligned training corpus is initially read, alignments for new sentences just fly out, so you do not need to wait several hours for Giza++ again. Also, unlike the above method, your original training alignment is now guaranteed to remain untouched by the new evidence.

Once all the input files are ready, the indexer is invoked as follows:

addictindex.pl \
    -trs train.en -trt train.hi -tra train.ali \
    -s test.en -r test.hi -h test.system.hi -ra test.ali -ha test.system.ali \
    -o $CGI

The indexer will copy the input files and output all index files into the $CGI folder where the CGI scripts will find them.

The error classifier currently uses its own monlingual word-alignment of reference translation and the hypothesis. It is invoked as follows:

${ADDICTER}/prepare/detecter.pl -s srcfile -r reffile -h hypfile [-a alignment] -w workdir

and it creates the files workdir/tcali.txt and workdir/tcerr.txt. The input files (src, ref and hyp) can also be gzipped. Custom alignment between hypothesis and reference can be supplied. If it is not supplied, then the default aligner (${ADDICTER}/testchamber/align-greedy.pl) is invoked.

Place the files tcali.txt and tcerr.txt in the experiment subfolder of $CGI and the error classes will be displayed during test data browsing in the viewer. The viewer can work with several alternating alignments (perhaps using different aligning algorithms) of the same data. For each of those alignments, you have to run detecter.pl separately.

First make sure that your web server is running and configured properly and that your index and data files have been prepared in the correct place. If you do not use your own web server, invoke the script server.pl in the main Addicter folder. It will say something like

Please contact me at: <URL:http://localhost:2588/cgi/index.pl>

which is the URL you should point your browser to. The server uses a randomly picked port number unless you specify it as a commandline parameter: server.pl 8080.

In the browser, you will see a list of experiments (all subfolders of $CGI). Start browsing your data by clicking on an experiment.

This research has been supported by the grant of the Czech Ministry of Education no. MSM0021620838 (2010), by the grants of the Czech Science Foundation no. P406/11/1499 and P406/10/P259, the Estonian Science Foundation target financed theme SF0180078s08 (2011) and by the project EuroMatrixPlus (FP7-ICT-2007-3-231720 of the EU and 7E09003+7E11051 of the Ministry of Education, Youth and Sports of the Czech Republic; 2011-2012).

• Mark Fishel, Ondřej Bojar, Daniel Zeman, Jan Berka: Automatic Translation Error Analysis. In: TSD 2011, Plzeň, Czechia, 2011
• Daniel Zeman, Mark Fishel, Jan Berka, Ondřej Bojar: Addicter: What Is Wrong with My Translations? In: The Prague Bulletin of Mathematical Linguistics, vol. 96, pp. 79–88; MT Marathon, Trento, Italy, 2011
• Maja Popović: Hjerson: An Open Source Tool for Automatic Error Classification of Machine Translation Output. In: The Prague Bulletin of Mathematical Linguistics, vol. 96, pp. 59–68; MT Marathon, Trento, Italy, 2011
• Jan Berka, Ondřej Bojar, Mark Fishel, Maja Popović, Daniel Zeman: Automatic MT Error Analysis: Hjerson Helping Addicter. In: Proceedings of LREC 2012, İstanbul, Turkey, 2012