The manual gives instructions on installing and running Depfix, our open-source tool for automatic post-editing of machine translation. We cover the steps required to use Depfix to process your own data. The manual also contains a description of the Depfix pipeline, including instructions that will enable you to modify the operation of Depfix.
Depfix is a tool for automatic post-editing of outputs of English-to-Czech machine translation (MT), especially phrase-based statistical machine translation (PB-SMT). Depfix uses a pipeline of natural language processing (NLP) tools to analyze the input sentences, such as taggers and parsers. The corrections Depfix performs are mainly rule-based, and rely heavily on the analyses provided by the tools.
Depfix was introduced in , and subsequent improvements were described especially in  and . In , Depfix has been released as an open source tool, under the GNU GPL v2 licence. For a comprehensive description of the whole Depfix system, please refer to .
In this manual, we give detailed instructions on installing and running Depfix, including a documentation of all commands and settings. We also give a general overview of the structure of the source codes, which may serve as a starting point for modifying them, or for using them for inspiration in your own post-editing system. Brief instructions are also bundled directly with Depfix in a README file.
- 1 System requirements
- 2 Installing Depfix
3 Running Depfix
- 3.1 The basics
- 3.2 Running the full version of Depfix
- 3.3 Depfix commands
- 3.4 Depfix variables
- 4 Delving deeper
- 5 Support
1 System requirements
You need to have at least 3.5 GB of RAM to run the default version of Depfix. So far, Depfix has only been tested on Linux. In principle, it should be possible to run Depfix on other operating systems as well, such as Microsoft Windows with Cygwin11https://www.cygwin.com/ – however, this has not been tested, and we can provide little or no support in that area.
2 Installing Depfix
|wget -O install.tgz http://goo.gl/FpIOuw|
|tar -zxvf install.tgz|
Please note that running the installation script will take tens of minuted or even hours, and some of the steps may occasionally ask you some questions – so you should leave the script running but check it time to time.
If this works for you: congratulations, you can now skip to a test run of Depfix – see Section 2.9 (you may try to do the test run even if the installation prints out error messages, as these may be non-fatal). Otherwise, you may want to inspect the error messages you get and try to find a solution yourself, or you may refer to the following sections.
Depfix is implemented in Treex framework and is contained in the Treex Subversion repository,22https://public:email@example.com/projects/tectomt_devel/ in the trunk/treex/devel/depfix directory. Thus, to install Depfix, please refer to the Treex installation manual at http://ufal.mff.cuni.cz/treex/install.html and follow the steps 1 (Perl), 2 (Treex::Core), 3 (Treex::EN), and 5 (SVN). Step 4 (TrEd) is optional; installing TrEd will allow you to inspect intermediate Depfix files that contain linguistic analyses of the sentences, but is not required if you are happy with only using the input and output, which is plaintext. Step 6 (Featurama) is not needed for Depfix.
There are several additional steps that are typically required when installing Treex. For ease of use, we list here the whole procedure that currently seems to work for Ubuntu 14.04. Still, please note that installing Treex is a complex task, and you may need to perform the steps differently on your machine. Also, some steps may not be necessary in your situation, such as installing packages that you have already installed.
All of the codes listed in the following section can be downloaded as bash scripts from http://ufallab.ms.mff.cuni.cz/~rosa/depfix/install.tar.gz, e.g. in the following way:
|wget -O install.tgz http://goo.gl/FpIOuw|
|tar -zxvf install.tgz|
Using the downloaded scripts may be more reliable then copy-pasting the codes from the PDF file.
2.3 Installing missing packages
We recommend installing several software packages before installing Treex.
|sudo apt-get install subversion libxml2-dev zlib1g-dev g++ \|
2.4 Installing Treex
This is just a very short version of the full Treex installation manual. If this works for you, then fine. If it doesn’t, please refer to the original manual at http://ufal.mff.cuni.cz/treex/install.html.
|# Prepare Perl environment|
|wget -O- http://cpanmin.us | \|
|perl - -l ~/perl5 App::cpanminus local::lib|
|eval `perl -I ~/perl5/lib/perl5 -Mlocal::lib`|
|echo '## Treex installation ##' >> ~/.bashrc|
|echo 'eval `perl -I ~/perl5/lib/perl5 -Mlocal::lib`' >> ~/.bashrc|
|grep bashrc ~/.bash_profile || echo 'source ~/.bashrc' \|
|# Install part of Treex that is on CPAN|
|cpanm --force Treex::Core Treex::EN|
|# Check out current version of full Treex from its SVN repository|
|# Password is "public"|
|svn --username public co $SVN_TRUNK/treex ~/treex|
|svn --username public co $SVN_TRUNK/libs/other ~/treex/oldlib|
|echo export PATH="$HOME/treex/bin:$PATH" >> ~/.bashrc|
|echo export \|
|echo export TMT_ROOT=$HOME/.treex >> ~/.bashrc|
2.5 Installing missing CPAN modules
We recommend installing several CPAN modules:
|cpanm --force forks Lingua::Interset POE Ufal::MorphoDiTa \|
|App::whichpm Class::Std Locale::Language String::Util \|
The list of the modules may change as Treex is further developed, so please install any other missing modules. Whenever you get an error message such as Can't locate Ufal/MorphoDiTa.pm in @INC (followed by many lines of garbage you can ignore), use the cpanm installer in the following way to obtain it:
In some cases, you may have to use the --force switch to overcome some errors – for example, some of the packages seems to have errors in tests, which leads to errors even if the package installs successfully.
2.6 Installing Treex models
Many parts of Depfix require model files to run. There is a set of models that are required in the current version and seem not to be downloaded automatically.33 Treex supports automatic download of missing files, but it may fail for various reasons. Thus, we recommend downloading them beforehand:
|for m in \|
|wget -O - \|
|| install -D /dev/stdin ~/.treex/share/data/models/$m|
Whenever Treex dies because it cannot find a required model file, it will report the path to the missing file; the error message will look something like this:
|Cannot find ~/.treex/share/data/models/some/model|
In such case, please download the missing file from:
Probably the easiest way to do that is using wget:
|wget -O - \|
|| install -D /dev/stdin ~/.treex/share/data/models/some/model|
2.7 Installing MGiza++
You may need to install the MGiza++ word-aligner if this is indicated by error messages. To do so, you should visit the webpage of MGiza++, http://sourceforge.net/projects/mgizapp/, and follow the instructions. Currently, the easiest way to install MGiza++ seems to be the following:
|# Use any directory you wish (and have write access to).|
|svn checkout http://svn.code.sf.net/p/mgizapp/code/trunk $MGIZA_DIR|
Next, you need to ensure that Treex finds your MGiza++ installation:
|mkdir -p ~/.treex/share/installed_tools/mgizapp|
|ln -s $MGIZA_DIR/mgizapp/inst \|
When running Depfix, it may happen that you will get an error message that looks like this:
|~/.treex/share/installed_tools/mgizapp/install contains binaries|
|compiled for 'x86_64-linux' but you are using|
Most likely, if you edit the Treex wrapper for MGiza++ (Treex::Block::Align::A::MGiza) and disable the architecture check, MGiza++ will run just fine. To do so, open ~/treex/lib/Treex/Block/Align/A/AlignMGiza.pm in a text editor (such as vim, emacs or gedit), and delete the line which looks like this (in current version it is line 34):
2.8 Installing CzechMorpho
You may need to install CzechMorpho if this is indicated by error messages. This is a Perl module which is not present on CPAN, so you will have to run the following commands to install it:
|svn export $SVN_TRUNK/libs/packaged/CzechMorpho/ /tmp/CzechMorpho/|
2.9 Testing run of Depfix
Once you have everything installed, go to the treex/devel/depfix directory of the SVN checkout using the command-line:
Depfix is operated using make targets (which are called commands in this document). To test if you have installed Depfix correctly and to download several required models, invoke the test command:
This will try to run Depfix on 1 testing sentence and will take about 5 minutes to run, plus time to download required files (several hundreds of MB).
When it finishes, the end of the standard output should look approximately like that (the lines have been shortened here to fit on page):
|DEPFIX [2014-08-27_20-44-21_2952406743_small_test]: eval_show ...|
|chgd NIST_0 NIST_1 diff BLEU_0 BLEU_1 diff ...|
|1 2.5483 3.2433 69.49 0.3349 0.4625 12.76 ...|
|make: Leaving directory `/home/rur/treex/devel/depfix'|
If the test ends successfully, printing out BLEU and other scores, you have Depfix installed and running. If not, please refer to the preceding steps, the Treex installation manual, your common sense, the Internet, and, if nothing helps, feel free to contact us (see Section 5).
3 Running Depfix
3.1 The basics
You can run the whole Depfix pipeline by invoking the default target:
This is equivalent to invoking all the commands corresponding to main Depfix processing steps:
|make init totreex tag run_giza ner_en parse \|
|fix deepfix write_sentences eval|
Without any settings, this is very similar to running make test, as the default input to Depfix is the 1 testing sentence.
The init command creates a new experiment directory and copies the input files into it. It is therefore wise to specify several settings when invoking it (Makefile variables are used for this):
the label to use for the new experiment directory
the source English data, i.e. the input of the SMT system
the Czech data to be processed, i.e. the output of the SMT system
the reference Czech translation to be used for evaluation44Currently, Depfix requires that REFERENCE_CS is provided, so if you do not have the reference translation, fill in the same value as for DATA_CS.
The format of the input data is plaintext in UTF-8 encoding, one sentence per line. Depfix expects the sentences to be in “human” format (e.g. not lowercased or tokenized).
Depfix comes with a sample WMT10 dataset, so you can run the following command to initialize a new experiment directory with that dataset:
|make init DIRLABEL=wmt10_test DATA_CS=data/wmt10_bojar-2012.cs \|
The output of the init command will look something like this:
|DEPFIX [2014-09-08_11-38-38_1255015118_wmt10_test]: init|
A new working directory is generated for the experiment (and the input files are copied into it). The name of the directory is printed to the standard output, as shown in the example above.55The experiment directory name format is y-m-d_h-m-s_random_DIRLABEL.
The next step is running Depfix on this directory. As you have already invoked the init target, you now need to run all but the first step. It is of course possible to list the steps manually, but an easier way is to use the default_ni target, which is identical to the default target except for not containing the init step (its name stands for “default, no init”).
Whenever you are invoking a target which does not create a new experiment directory, but uses an already existing experiment directory, you must always specify the experiment directory using the DIRNAME setting.
Thus, to run Depfix on the experiment directory created in the previous step, it suffices to use:
|make default_ni DIRNAME=<<\direx|
The sample WMT10 dataset has 2489 sentences; expect Depfix to run for about 2 hours on it. Once it is finished, it will print our automatic evaluation results and exit.
Depfix output can be found in the output.txt file in the experiment directory, or displayed using the view target:
|make view DIRNAME=<<\direx|
To see what has been changed by Depfix, use the compare target (or compare_log to also see details about the fixes applied):
|make compare DIRNAME=<<\direx|
There are many other commands available, although the already mentioned ones are sufficient for basic operation of Depfix. For a full documentation, see Section 3.3.
There are also many settings available, most of them specifying which Treex scenario should be used for each processing step. Two more are detailed in the next section; the other can usually be left at their default values. For a full documentation, see Section 3.4.
3.2 Running the full version of Depfix
By default, only a basic version of Depfix is run. The full state-of-the-art version differs from the basic version in several aspects:
it uses a parser adapted to parsing SMT outputs
it uses a statistical fixing component
it requires about 20 GB of RAM to run
it runs about twice as long as the basic version
it achieves slightly better results – e.g. on the sample WMT10 data, applying the full version leads to an improvement of 0.45 BLEU point, while the basic version achieves a 0.37 BLEU point improvement
The parser used by default for Czech is a version of the Maximum Spanning Tree (MST) parser , adapted for standard Czech , but not for SMT outputs. If you have more than 20 GB of RAM, you should use the MSTperl parser adapted for SMT outputs [4, 9] instead, which is more robust to various grammatical errors commonly occurring in SMT outputs, and uses additional features that have shown to be useful in that setting (such as features based on the source English sentence parse tree). To switch to this parser, use the following additional setting when running Depfix (more specifically, when invoking the parse target):
You can also enable the statistical fixing component , which targets errors in valency of verbs and nouns. It requires at least 15 GB RAM to run. We do use it in our setups, but please note that evaluation of the effect of using the component has brought inconclusive results – it usually improves the results, but only very slightly. To enable the component, use the following setting (relevant for the deepfix target):
If you want to use full Depfix by default, you can make these (as well as any other) settings default by using the settings.mak file; instructions are included in the file itself (edit the file using a plaintext editor).
3.3 Depfix commands
Depfix commands are Makefile targets – you may inspect them by viewing the source code of depfix/Makefile. This section documents the current set of available commands. Please note that Depfix is an experimental software in active development, and commands may be added, removed or modified; also, some commands may become temporarily unusable.
Most of the commands are parametrized with variables, listed in the following section. In this section, we give the names of the variables with their default values wherever applicable – e.g. OUTPUT_TXT=output.txt refers to a variable called OUTPUT_TXT whose default value is output.txt.
As already mentioned, the most important variable is DIRNAME, which specifies the working directory to be used, and must be set for any command that does not create a new directory. Do not set DIRNAME if you invoke the default command, the init command, or the clone command. For any other target, you must set DIRNAME – e.g.:
|make view DIRNAME=<<\direx|
Most of other variables specify scenarios to use for the commands – their value is always a name of a file in the depfix/scenarios subdirectory – or filenames to use. You do not have to set these variables unless you want to modify the way Depfix behaves, their default values are just fine.
While running, Depfix prints brief information to the standard output, listing the command currently being run (including subcommands), the experiment directory, and some of the settings. More detailed information is printed to the standard error output.
3.3.1 Main Depfix pipeline
The main Depfix pipeline is invoked by the default command, which invokes a set of subcommands that correspond to individual steps of the processing pipeline. Invoking those commands explicitly is useful if you need to run only a part of the pipeline – for example, if you want to play with the fixing rules, you may run the analysis up to the parse step, make several copies of the experiment directory, and then run the rest of the pipeline on each of the directories separately with different fixing scenarios.
At end of the parse step, the intermediate files are backed-up in a parsed subdirectory of the working directory, and they are restored at beginning of the fix step. Thus, if you run the fix step on an already fixed directory, the intermediate files will first be restored to the state in which they were after the parse step. This is generally not true for the other steps, so running them more than once on one directory may lead to errors. Still, if a step dies, it is often safe to try to run it again – e.g. if the pipeline dies while performing the run_giza step, it is worth trying to rerun the pipeline again by explicitly listing all the steps starting at run_giza.
Run the whole Depfix pipeline on the input data (see init for important variables to set). Typing make without specifying any command will also invoke the default command. default invokes the following commands: init totreex tag run_giza ner_en parse fix deepfix write_sentences eval
Identical to default except for not invoking the init command. The DIRNAME variable must be specified. To be used if you prefer to init your working directory in a separate step, and then run Depfix on it later.66This is quite a common practice – you may want to prepare your experiment directories by running init fully manually, as this involves setting a lot of variables and takes very little time to run, and then let run default_ni without paying attention to it in parallel for several experiments, as this can be fully automated and usually takes hours.
Create a new experiment directory and copy input data into it. Set a name for the new directory using DIRLABEL.77 The experiment directory name format is y-m-d_h-m-s_rand_DIRLABEL and the name of the newly created directory, together with other information, is printed to standard output. Specify the input data using DATA_CS, DATA_EN and REFERENCE_CS; the value of each of these variables must be a path to a text file (UTF-8, one sentence per line). The default values correspond to the sample 1-sentence input used in test (DATA_CS=data/csw.txt DATA_EN=data/en.txt REFERENCE_CS=data/cs.txt); to use a larger sample input, you may use the following setting:
DATA_CS=data/wmt10_bojar-2012.cs DATA_EN=data/wmt10_src.en REFERENCE_CS=data/wmt10_ref.cs
The reference Czech translation is only used in the last step – it is not analyzed by NLP tools and is not used in the fixing.
Read the input text data into Treex files and perform tokenization projection. The resulting files are called translation001.treex.gz, translation002.treex.gz, etc., and each contains 100 sentences (or, more precisely, 100 source English sentences + 100 Czech machine translation sentences + 100 Czech reference translation sentences). The following steps operate on these files – each step reads them from disk when it starts and saves them modified back to disk when it ends. The files are in PML, which is an XML format, and can be viewed by the TrEd editor. The tokenization projection is performed by the scenario specified in PROJECT_TOKENIZATION_SCEN=project_tokenization.scen.
Tokenize the sentences and run morphological analyses on them. (EN_ANALYSIS_1_SCEN=en_analysis_1.scen, CS_ANALYSIS_1_SCEN=cs_analysis_1_morphodita.scen). After that step, each sentence is represented by a set of nodes, corresponding to the tokens, and each of the nodes has the following three attributes filled: form, lemma, and tag. The tokens are contained in an a-tree (although the tree is flat at this moment – all nodes are children of the technical root node).
Word-align the sentences (RUN_GIZA_SCEN=run_mgiza.scen) and run heuristics to add missing alignment links (ADD_MISSING_LINKS_SCEN=add_missing_links.scen). The word-alignment used is of the intersection type, i.e. for each token there is at most one aligned token. The alignment links are stored in attributes of the English nodes.88However, this is not very practical for operation of Depfix, which operates on the Czech nodes. Therefore, the alignment will be stored in attributes of the Czech nodes in near future.
Run named entity recognizer for English (NER_EN_SCEN=ner_en.scen). This creates an n-tree which contains named entities found in the sentence.
Run a dependency parser on the sentences (EN_ANALYSIS_2_SCEN=en_analysis_2.scen, CS_ANALYSIS_2_SCEN=cs_analysis_2_msta.scen). After this step, the a-trees created in the tag step are no more flat but instead represent the dependency structure of the sentence. This step consists of two substeps: parse_only, which does the parsing, and parse_backup, which copies the treex files into the parsed subdirectory.
Run the Depfix fix rules. This step consists of four substeps. First, restore_parsed replaces the current treex files by the files from the parsed subdirectory. Next, fix_prepare prepares the files for Depfix by copying the a-trees into a new T zone, so that the files will contain both the original and the fixed sentences for easy observation of the fixes performed (FIX_PREPARE_SCEN=fix_prepare.scen). After that, ner_cs runs a named entity recognizer for Czech (NER_CS_SCEN=ner_cs.scen). And finally, fix_run runs the actual fix rules (FIX_SCEN=fix.scen).
Run the fixes that operate on t-layer. This command consists of four subcommands. a2t creates a t-layer analysis (A2T_CS_SCEN=a2t_cs.scen, A2T_EN_SCEN=a2t_en.scen). tfix_prepare adds word alignment links from a-layer (TFIX_PREPARE_SCEN=tfix_prepare.scen). tfix performs the t-layer fixes (TFIX_SCEN=tfix_rules.scen); the fixes are immediately projected to a-layer. refix_after_tfix runs the a-layer fixes again especially in order to fix agreement that may have been violated while performing t-fixes (REFIX_SCEN, by default identical to FIX_SCEN).
Write the fixed sentences into OUTPUT_TXT=output.txt (WRITE_SENTENCES_SCEN=write_sentences.scen). Also calls a write_fixlog substep that writes a log containing the fixes performed into fixlog.txt.
Perform automatic evaluation, storing the result into AUTOEVAL_OUT=autoeval.out. This step consists of several substeps, each of which computes a different characteristic and stores it. You may specify a different reference translation file to be used (REF_CS_TXT=ref_cs.txt). eval_dirname stores the name of the working directory. eval_lines computes the number of sentences modified by Depfix. eval_bleu computes BLEU and NIST scores. eval_ter computes PER and TER scores. eval_show processes the contents of the AUTOEVAL_OUT file and prints out the results formatted as tab separated values.
3.3.2 Viewing the results of Depfix run
Most of the commands in this section can be parametrized using the following variables, specifying the names of files in the experiment directory to use (but typically the default values, given in brackets, are what you want, so you usually do not need to set these):
Original Czech translation (data_cs.txt)
Depfix output (OUTPUT_TXT)
Reference Czech translation (REF_CS_TXT)
Source English sentences (data_en.txt)
The names of these variables are also used as labels for the sentences in the compare* commands.
The following commands are available for displaying the results of Depfix in various ways:
Show the output of Depfix, stored in file specified by OUTPUT_TXT. Only the resulting sentences are show, one sentence per line.
Show the output of Depfix together with the inputs. For each sentence, the original Czech sentence (ORI) and the fixed sentence (NEW) are compared, the differences are highlighted, and the sentences are printed together with the source English sentence (EN) and the reference Czech translation (REF). If there is no difference between the original and the fixed sentence, the sentences are not printed.
Similar to compare, but also showing a fix log, which contains a list of fixes that were applied on the sentence.
Similar to compare, but the search for differences is done case-insensitively – differences that are only in casing are not highlighted, and sentences in which Depfix changed only casing are not considered as changed (and therefore not printed).
Similar to compare_log, but the search for differences is done case-insensitively as in compare_ci.
Similar to compare, but the output is in HTML format (differences are highlighted by bold font). The resulting HTML file is printed to standard output – use stdout redirection to save it into an HTML file.
Compare results of two experiments run on the same input data – differences between the two outputs of Depfix are highlighted, instead of differences between the input and output. The two experiment directories are to be specified by DIRNAME and DIRNAME2:
make compare2 \ DIRNAME=2014-07-15_17-37-21_4908631573_wmt10_test \ DIRNAME2=<<\direx
Similar to compare2, but the output is in HTML format as in comparehtml.
3.3.3 Manual evaluation
There is a set of commands for carrying out manual evaluation of Depfix processing (maneval_prepare, maneval_eval, cross_annot_agree, cross_annot_agree_matrix, and other). As the needs for manual evaluation are typically different every time the evaluation is performed, we suggest you use these commands as inspiration when performing manual evaluation – please inspect the source code of the Makefile for the code used.
The commands make use of a simple manual evaluation tool by Ondřej Bojar, called QuickJudge.99http://ufal.mff.cuni.cz/euromatrix/quickjudge/ The tool is bundled with Depfix, as it consists of one Perl script, and resides in depfix/scripts/quickjudge.pl. It generates text files with randomized order of outputs of various MT systems or setups, which makes blind evaluation possible. When annotators have marked in the text files which translations are better, the tool then processes the annotations, which makes it possible to compute evaluation results.
Still, if you need to perform a manual evaluation of Depfix and do not want to bother with tweaking the commands, the basic procedure that works out-of-the-box is the following:
Run the maneval_prepare command. You must set DIRNAME, and you may set other variables: the number of lines to be annotated (ANNOT_LINES=3003), the number of chunks to split these lines into (ANNOT_NUM=2), and the prefix to use for the newly created manual evaluation directory (MAN_PREFIX=depfix_maneval). So, the command you run may look like that if you want to evaluate 500 sentences by 5 annotators:
make maneval_prepare ANNOT_LINES=500 ANNOT_NUM=5 \ DIRNAME=<<\direx
The manual evaluation directory will have a name following the pattern MAN_PREFIX_DIRNAME and will contain a number of generated files. Please note that the maneval_prepare command splits the files before looking at the changes, so if you e.g. want to evaluate the first 500 sentences (ANNOT_LINES=50), you may expect to obtain about 300 sentences for annotation, as many sentences are not changed by Depfix and there is nothing to annotate on them. For the same reason, if you split these into 5 chunks, (ANNOT_NUM=5), each chunk will contain a different number of sentences (e.g. the third chunk will contain sentences from 201st to 300th that were changed by Depfix).
Take the *.anot files from the manual evaluation directory and give them to your annotators with appropriate instructions. Files are to be annotated by adding any one character (e.g. 1, *, y…) to the beginning of the line which is BETTER than the other line, e.g.:
EN_ORIG A big role in the film is played by Matt Damon. CS_REF Velkou roli roli hraje herec Matt Damon. * Velkou roli ve filmu hraje Matt Damon. Velkou roli ve filmu hrajou Matt Damon.
Do not do anything with the line which is not the good one. If quality of the lines is equal, do not add anything anywhere, just leave it as it is. The lines start with a TAB character, which must be kept. The lines marked EN_ORIG and CS_REF are just for information. Do not add any newlines!
Put the annotated files back into the manual evaluation directory (overwrite their unannotated versions, these are not needed any more), and run maneval_eval (you need to specify DIRNAME; if you used a custom MAN_PREFIX, you need to specify that as well).
3.3.4 Other commands
Run the whole Depfix pipeline on a sample input, consisting of only one sentence. You do not need to set any variables for this command, everything is already preset.
Display the README file, which contains brief instructions on running Depfix.
Copy an experiment directory (OLDDIRNAME) to a new one (created according to DIRLABEL). Similar to performing cp -r, but including the generation of a new DIRNAME as in init, i.e. containing the date and time. As the clone command fills the name of the generated directory into the DIRNAME variable, it can be followed by other Depfix commands.
make clone DIRLABEL=experiment_2 \ OLDDIRNAME=<<\direx
Run bootstrap resampling significance test to estimate whether the difference in BLEU score achieved by Depfix post-editing (as computed by eval) is statistically significant. The number of samples is set by SAMPLES=100, the significance level by ALPHA=0.05. Use with caution, as some authors do not regard the results of bootstrap resampling as credible.
Several technical or experimental commands are not described in this manual.
3.4 Depfix variables
Depfix commands can be parametrized by using variables (these are standard Makefile variables). All of the variables are set to default values, so in most cases you will not need to set them explicitly. However, there are several variables that must be always specified:
the variables for the init target, which specify the input files and the label for the new experiment directory (DATA_CS, DATA_EN, REFERENCE_CS, DIRLABEL),
the OLDDIRNAME variable for the clone target,
and, most importantly, the DIRNAME variable, which specifies the experiment directory to be used, and must be specified for any set of commands that does not start with one of the following commands: init, default, clone.
Several technical or experimental variables are not described in this manual – we focus on variables that may be worth changing in some situations.
3.4.1 Setting the variable values
Depfix variables may be set on the command line when invoking a Depfix command:
|make init DIRLABEL=wmt14_bojar|
Moreover, the default settings of the variables can be overridden by editing the settings.mak file in a plaintext editor. The file is loaded after the Makefile, so settings in settings.mak have priority over the defaults specified in the Makefile; however, you can still override these by setting the variables on the command line, which has always the highest priority.
The settings.mak file contains information on how to use it; generally, you set a new default value for a variable in the following way – e.g. to use a different scenario for writing out the sentences without performing detokenization:
If a setting appears multiple times in the file, the last occurrence is the one that is valid.
The file also contains prepared settings for running the full version of Depfix (see Section 3.2), and for running Treex on an SGE cluster (these settings are commented out by default).
3.4.2 Input and output files
The specification of input files is detailed in the decryption of the init command in Section 3.3.1 – DATA_EN is the source English text (input of an MT system), DATA_CS is its Czech machine translation (output of the MT system), and REFERENCE_CS is a reference Czech translation of the source English text by a human translator.
The default values correspond to the sample one-sentence input used for the test command:
|DATA_CS=data/csw.txt DATA_EN=data/en.txt REFERENCE_CS=data/cs.txt|
You can also use the larger sample input which is bundled with Depfix:
The format of the input files must be plain text in UTF-8 encoding, one sentence per line. All of the input files must have the same number of lines. Currently, Depfix requires you to provide a reference translation file; if you do not have one, please set REFERENCE_CS to any other file (which has the correct number of lines); we recommend setting to the same value as DATA_CS.
The input files get copied into the experiment directory under the following filenames: data_cs.txt, data_en.txt, and ref_cs.txt. The output of the Depfix pipeline is stored into a text file in the experiment directory, specified by OUTPUT_TXT (the default is output.txt); the log which contains information on fixes that were applied is stored in fixlog.txt.
The compare* commands (see Section 3.3.2) operate on the inputs and outputs of Depfix by default, but can be parametrized to behave differently by setting the following variables:
original Czech translation (defaults to data_cs.txt)
Depfix output (defaults to OUTPUT_TXT)
reference Czech translation (defaults to REF_CS_TXT, which defaults to ref_cs.txt)
source English sentences (defaults to data_en.txt)
The output of automatic evaluation is stored into a text file specified by AUTOEVAL_OUT; the default is autoeval.out.
3.4.3 Depfix operation
is the label to use for the new directory created by init (or default, or clone).
is the command to invoke Treex (defaults to treex); you may add some Treex options in special cases
is the command to invoke Treex in parallel, i.e. to perform commands that can be run independently on individual Treex files (generally, all commands except those that work with text files – these use TREEX instead of TREEXP). It defaults to treex, which means that Treex is never run in parallel by default. If you have an SGE cluster, you may use settings such as the following to run Treex in parallel by default (it is reasonable to set these in the settings.mak file, as described in Section 3.4.1):
JOBS=10 MEM=30G WORKDIR=$$(mktemp -d --tmpdir=$(DIRNAME)) TREEXP=treex -p --survive --jobs=$(JOBS) --mem=$(MEM) \ --workdir=$(WORKDIR)
Most of the settings specify scenarios to use for various commands, as described in the previous section. The scenarios themselves are contained in the depfix/scenarios subdirectory. In many cases, there are alternative scenarios available that you may use in some situations to modify the way Depfix operates, such as to use a different parser or to omit some processing steps; the name of an alternative scenario typically has an identical prefix to that of the default scenario.
Also, there is the empty.scen scenario, which does not do anything, so you may use it to bypass some scenario completely (although this will fail in some cases, as many of the steps are vital for the operation of the subsequent steps). For example, to skip running Czech named entity recognizer, you may use the following setting:
We list here the settings, their default values, and some of the alternatives in cases where we find it useful. The names of the settings are rather informative, and they are given in the order in which they are used in the default Depfix pipeline; to see exactly where they are used, see Section 3.3.
Default value: project_tokenization.scen
Default value: en_analysis_1.scen. Alternative value: en_analysis_1_old.scen to use the Morče tagger
Default value: cs_analysis_1_morphodita.scen. Alternative values: cs_analysis_1_featurama_fme.scen to use the Featurama tagger, or cs_analysis_1_morce.scen to use the Morče tagger
Default value: run_mgiza.scen. Alternative value: run_mgiza_ufal.scen to be used on machines of our institute (ÚFAL); the important difference is tmp_dir=/COMP.TMP instead of tmp_dir=/tmp – if your /tmp directory is on a network drive, you should use a similar modification so that a local directory is used instead
Default value: add_missing_links.scen
Default value: ner_en.scen
Default value: en_analysis_2.scen
Default value: cs_analysis_2_msta.scen. Alternative value: cs_analysis_2_boost_model_025.scen to use the MSTperl parser adapted for parsing SMT outputs (see Section 3.2)
Default value: ner_cs.scen
Default value: fix_prepare.scen
Default value: ner_cs.scen
Default value: empty. May be used to run a scenario before running fix.scen in the fix_run command
Default value: fix.scen
Default value: a2t_cs.scen
Default value: a2t_en.scen
Default value: tfix_prepare.scen
Default value: tfix_rules.scen. Alternative values: tfix_cut_ChgCase2.scen to also run statistical fixes (see Section 3.2); tfix_stat.scen to run only the statistical fixes, i.e. not running the rules based ones.
Default value: $(FIX_SCEN)
Default value: write_sentences.scen. Alternative value: write_sentences_no_detok.scen not to perform detokenization, i.e. to write out individual tokens separated by spaces.
The following variables are used for manual evaluation (see Section 3.3.3):
The number of sentences to be selected for evaluation (including unchanged ones, so the final number of sentences is likely to be lower). Default value: 3003
Number of annotation files to generate – usually should be a multiple of the annotators you have, as each of the generated files is to be given to one of your annotators. Default value: 2
Prefix to use for the directory created for manual evaluation files. Default value: depfix_maneval
The following variables are used for bootstrap significance tests in automatic evaluation (see Section 3.3.4):
The number of repetitions of the experiment. Default value: 100
The significance level. Default value: 0.05
4 Delving deeper
If you want to modify Depfix operation more than what is allowed by changing the settings of the variables, you need to go into the source codes – the Makefile, the scenarios, and/or the Treex blocks.
One possibility you have is modifying the scenarios used by Depfix; we suggest that you do that by copying an existing scenario, modifying it, and then calling Depfix with the corresponding value set to the filename of such new scenario. All the scenarios are plaintext files that reside in the depfix/scenarios subfolder; do put even your new scenarios in that folder for smooth operarion. A scenario typically sets several global arguments, and then invokes a pipeline of Depfix blocks, sometimes with arguments set. See the Treex manual for more information about Treex scenarios.
You may also need to edit the Makefile to suit your modified scenarios; see a make manual if you need to.
The key scenario which invokes most of the fixes is fix.scen. The fixes are Treex blocks, and thus can be found in subdirectories of treex/lib/Treex/Block; most of them are in the treex/lib/Treex/Block/A2A/CS directory (but not everything in this directory are Depfix blocks; Depfix block names typically start with the word Fix).
There are also several Depfix-specific Treex tools used; you will find them in treex/lib/Treex/Tool/Depfix. These are mainly wrappers for existing tools to enable easy operation from within Depfix.
Depfix is supported by the grants FP7-ICT-2013-10-610516 (QTLeap), GAUK 1572314, and SVV 260 104. This work has been using language resources developed, stored and distributed by the LINDAT/CLARIN project of the Ministry of Education, Youth and Sports of the Czech Republic (project LM2010013).
-  (2011) Two-step translation with grammatical post-processing. Edinburgh, UK, pp. 426–432. Cited by: Introduction.
-  (2005) Non-projective dependency parsing using spanning tree algorithms. Association for Computational Linguistics , Vancouver, British Columbia, Canada, pp. 523–530. Cited by: 3.2.
-  (2007) Feature engineering in maximum spanning tree dependency parser. Lecture Notes in Computer Science, Pilsen, Czech Republic, pp. 92–98. External Links: Cited by: 3.2.
-  (2012) Using parallel features in parsing of machine-translated sentences for correction of grammatical errors. Jeju, Korea, pp. 39–48. Cited by: 3.2.
-  (2012) DEPFIX: a system for automatic correction of Czech MT outputs. Montréal, Canada, pp. 362–368. Cited by: Introduction.
-  (2013) Deepfix: statistical post-editing of statistical machine translation using deep syntactic analysis. Sofija, Bulgaria, pp. 172–179. Cited by: 3.2, Introduction.
-  (2013) Automatic post-editing of phrase-based machine translation outputs. Master’s Thesis, Charles University in Prague, Faculty of Mathematics and Physics, Praha, Czechia. External Links: Cited by: Introduction.
-  (2014) Depfix, a tool for automatic rule-based post-editing of SMT. The Prague Bulletin of Mathematical Linguistics 102, pp. 47–56. External Links: Cited by: Introduction.
-  (2014) MSTperl parser. Charles University in Prague, UFAL (ces, eng). Note: http://hdl.handle.net/11858/00-097C-0000-0023-7AEB-4 Cited by: 3.2.