bitextor
is a tool to automatically harvest bitexts from multilingual websites. To run it, it is necessary to provide:
- The source where the parallel data will be searched: one or more websites (namely, bitextor needs website hostnames)
- The two languages on which the user is interested: language IDs must be provided following the ISO 639-1
- A source of bilingual information between these two languages: either a bilingual lexicon (such as those available at the bitextor-data repository), a machine translation (MT) system, or a parallel corpus to be used to produce either a lexicon or an MT system (depending on the alignment strategy chosen, see below)
Apart from downloading all submodules of this repository (you can do it with git clone --recurse-submodules https://github.com/bitextor/bitextor.git
if you are cloning this repo from scratch or in case you are downloading a tarball just do git submodule update --init --recursive
), there are some external tools that need to be in the path before installing the project. autotools and pkg-config are necessary for building and installing the project. Tools from JDK are needed to run Java dependences (Boilerpipe); version 8 or later are required. In addition, a c++ compiler is required for compiling dependences. libboost-all-dev dependence is need to compile clustercat
and mgiza
projects. Optionally, httrack can be used for crawling if it is installed.
If you are using an apt-like package manager you can run the following command line to install all these dependences:
sudo apt install cmake automake pkg-config python3 python3-pip libboost-all-dev openjdk-8-jdk liblzma-dev
Furthermore, most of the scripts in bitextor are written in Python 3. Because of this, it is necessary to install Python >= 3. All these explained tools are available in most Unix-based operating systems repositories.
Some additional Python libraries are required. They can be installed automatically with the tool pip by runing (use without sudo
if you are running in a virtualenv):
pip3 install -r requirements.txt
pip3 install -r bicleaner/requirements.txt https://github.com/vitaka/kenlm/archive/master.zip
As we explained above, the web crawler HTTrack can be used in Bitextor. To do so, first install it by running the command:
sudo apt install httrack
This dependency is not mandatory as a second parallel data crawler is provided in Bitextor (Creepy).
To compile all bitextor submodules you will first need to run the script configure
(if you are downloading the code directly from the repository you will need to run the script autogen.sh
instead, which will identify the location of the external tools used). Then the code will be compiled using make
:
./autogen.sh && make
In some machines equipped with an AMD CPU you may experience some troubles with tensorflow 1.8.0 (the version specified in requirements.txt). In case you have installed all the requirements successfully, but when running ./autoconf.sh or ./configure you get an error that says tensorflow is not installed, please, replace current version with version 1.5:
sudo pip3 uninstall tensorflow
sudo pip3 install tensorflow==1.5.0
In addition, some users haver reported problems when trying to install tensorflow using pip3
for versions of Python >= 3.7. If this is the case, you can try to install it manually or using another package management tool, or to use a lower version of Python.
Depending on the version of libboost that you are using, you may experience some problems when compiling some of the sub-modules included in Bitextor. If this is the case you can install version 1.66 manually by running the following commands:
sudo apt-get remove libboost-all-dev
sudo apt-get autoremove
wget https://dl.bintray.com/boostorg/release/1.66.0/source/boost_1_66_0.tar.gz
tar xvf boost_1_66_0.tar.gz
cd boost_1_66_0/
./bootstrap.sh
./b2 -j16 --layout=system install || echo FAILURE
cd ..
rm -rf boost_1_66_0*
To run Bitextor use the main script bitextor.sh
. In general, this script will take two parameters:
bitextor.sh -s <CONFIGFILE> [-j <NUMJOBS>]
where
<CONFIGFILE>
is a YAML configuration file containing the list of parameters to run bitextor (learn more about bitextor configuration in the next section), and<NUMJOBS>
is the number of jobs that can be launched in parallel; a job is a single step of the pipeline (see section Pipeline description) and can be run in parallel for different websites For example, on a machine with 4 cores, one could run Bitextor as follows:
bitextor.sh -s myconfig.yaml -j 4
If bitextor is run on a cluster with a software that allows to manage job queues, two more options can be used:
bitextor.sh -s <CONFIGFILE> [-j <NUMJOBS>] [-c <CLUSTERCOMMAND>] [-g <CLUSTERCONFIG>]
where
<NUMJOBS>
is redefined as the number of jobs that can be submitted to the cluster queue at the same time,<CLUSTERCOMMAND>
is the command that allows to submit a job to a cluster node (for example, this command would besbatch
in SLURM orqsub
in PBS),<CLUSTERCONFIG>
is a JSON configuration file that allows to specify the specific requirements for each job in the cluster (for example, this file allows to specify if a job requires more RAM memory, or GPUs available, for example). Further information about how to configure job requirements in a cluster can be obtained in Snakemake's documentation.
In the case of running on a cluster with, for example, the SLURM workload manager installed, one could run Bitextor as:
bitextor.sh -s myconfig.yaml -j 20 -c "sbatch"
this command would run bitextor allowing to submit 20 jobs in the cluster queue at the same time, assuming that all jobs can be run in any node of the cluster.
Now assume that we plan to train a neural MT (NMT) system with bitextor for document alignment (see next section). In this case, we would need to configure the call to the cluster in a way that those rules that require using GPUs for training or running NMT are run in nodes with GPUs. We could create a cluster configuration file such as the following (extracted from snakemake/examples/cluster.json
):
{
"__default__" :
{
"gres": "",
},
"docaling_translate_nmt" :
{
"gres": "--gres gpu:tesla:1"
},
"train_nmt_all":
{
"gres": "--gres gpu:tesla:1"
}
}
this configuration file is telling the cluster to set option gres
empty for all jobs but docalign_translate_nmt
and train_nmt_all
for which it would take value --gres gpu:tesla:1
. In SLURM --gres
is the option that allows to specify a resource when queuing a job; in the example we would be specifying that a tesla GPU is required by these two jobs. Once we had our configuration file, we could call bitextor in the following way:
bitextor.sh -s myconfig.yaml -j 20 -c "sbatch {cluster.gres}" -g cluster.json
Note that, in this case, an additional option needs to be added to the sbatch
command so it is called using the specific gres
option as indicated in the config file cluster.json
described above: it will be empty for most jobs but for docalign_translate_nmt
and train_nmt_all
.
Bitextor uses a configuration file to define the variables required by the pipeline. Depending on the options defined in this configuration file the pipeline can behave differently, running alternative tools and functionalities. The following is an exhaustive overview of all the options that can be set in the configuration file and how they affect to the pipeline.
Suggestion: A minimalist configuration file sample (default.yaml
) can be found in this repository (snakemake/example/tests/default.yaml
). You can take it as an starting point by changing all the paths to match your environment.
There are a few variables that are mandatory for running bitextor, independently of the task to be carried out:
permanentDir: /home/user/permanent/bitextor-output
transientDir: /home/user/transient
lang1: en
lang2: fr
permanentDir
andtransientDir
: Folders used during processing:permanentDir
will contain the results of crawling, i.e. the parallel corpus built and the WARC files obtained through crawling;transientDir
will contain the rest of files generated during processinglang1
andlang2
: Languages for which parallel data is crawled; note that if MT is used in the pipeline (either for alignment or evaluation) the translation direction used will be lang1 -> lang2
There are some more options that are rather basic but not mandatory as they take default values if they are not defined
bitextor: /home/user/bitextor
LANG1Tokenizer: /home/user/bitextor/preprocess/moses/tokenizer/tokenizer.perl -q -b -a -l en
LANG2Tokenizer: /home/user/bitextor/preprocess/moses/tokenizer/tokenizer.perl -q -b -a -l fr
LANG1SentenceSplitter: /home/user/bitextor/preprocess/moses/ems/support/split-sentences.perl -q -b -l en
LANG2SentenceSplitter: /home/user/bitextor/preprocess/moses/ems/support/split-sentences.perl -q -b -l fr
temp: /home/user/transient
maxSizeWARC: 1000
boilerpipeCleaning: true
alcazar: false
bitextor
: Directory where Bitextor is installed (the repository or tarball downloaded and compiled); if it is not specified it is assumed that it is the director where the scriptbitextor.sh
isLANG1Tokenizer
andLANG2Tokenizer
: scripts for word-tokenization both forlang1
andlang2
. These scripts must read from the standard input and write to the standard output. If no tokenizer is set the one provided by the Moses toolkit is used.LANG1SentenceSplitter
andLANG2SentenceSplitter
: scripts for sentence splitting both forlang1
andlang2
. Again the scripts must read from the standard input and write to the standard output. If no sentence splitter is set the one provided by the Moses toolkit is used.temp
: temporary directory where some files that will be only needed for a single job will be stored; if it is not defined it is set to the same directory astransientDir
maxSizeWARC
: when a website is crawled, all the documents downloaded are stored into a WARC file; this option allows to specify the maximum size of a WARC file, so when it is reached the WARC file is split into n files containing, as much, the maximum value set. This allows to run pre-processing in parallel for each of the WARC files obtained. Smaller values of this option implies a higher number of WARC files that can be pre-processed in parallel which, depending on the resources available, may result in a faster running of BitextorboilerpipeCleaning
: option that enables the use of the tool boilerpipe to remove boilerplates from HTML documents; by default this is disabledalcazar
: option that enables the library alcazar for text extraction from HTML documents; by defaultlxml
library is used
The next set of options refer to the source from which data will be crawled. Two options can be specified for crawling: one is to specify a list of websites to be crawled, while the other one is to provide a langstat file containing language statistics regarding the documents in one or more websites, so promising websites can be identified.
hosts: ["www.elenacaffe1863.com","vade-retro.fr"]
langstat: /home/user/langstat/langstats.all.gz
langstatThreshold: 50
langstatExcludeDomains: /home/user/bitextor/snakemake/exclude-domains
hosts
: list of hosts to be crawled; the host is the part of the URL of a website that identifies the web domain, this is, the URL without the protocol and the path. For example, in the case of the url https://github.com/bitextor/bitextor the host would be github.comlangstat
: file containing language statistics of a collection of websites (hosts). The langstat file is a tab-separated list of tuples host - language - amount of documents. For example:
0-0hamster.livejournal.com el 17
0-0hamster.livejournal.com en 1102
0-0hamster.livejournal.com hi 19
0-0hamster.livejournal.com ms 33
0-0hamster.livejournal.com nn 29
langstatThreshold
: minimum number of documents in each language so the web domain is considered for crawling.
Two crawlers are supported by Bitextor: one is based on the library Creepy and the other on the external tool HTTrack. The following are the variables that allow to choose one of them and to configure some aspects of the crawling.
httrack: true
crawlTimeLimit: 30s
crawlSizeLimit: 1G
crawlTld: false
crawlerNumThreads: 1
crawlerConnectionTimeout: 10
httack
: if this option is enabled, HTTrack is used instead of the crawler based on CreepycrawlerUserAgent
: user agent to be added to the header of the crawler when doing requests to a web server (identifies your crawler when downloading a website)crawlTimeLimit
: time (in seconds) for which a website can be crawled; for example: 3600s for a crawl of an hourcrawlSizeLimit
: creepy-specific option that limits the size of the crawl, i.e. when this limit is reached the crawl ends; it can be specified in GB (G), MB (M) or KB (K)crawlTld
: creepy-specific option that allows the crawler to jump to a different web domain as far as it is part of the same top-level domain (TLD); a TLD could be, for example, .es, .info or .orgcrawlerNumThreads
: creepy-specific option that allows to specify the number of threads to be be used by the crawler; by default this number is 1crawlerConnectionTimeout
: creepy-specific option that allows to specify the connection timeout to a web server
Two strategies are implemented in bitextor for document alignment. The first one uses bilingual lexica to compute word-overlapping-based similarity metrics; these metrics are combined with other features that are extracted from HTML files and used by a linear regressor to obtain a similarity score. The second one uses machine translation (MT) and a TF/IDF similarity metric computed on the original documents in lang1
and the translations of documents in lang2
. Bitextor allows to build (if necessary) both the bilingual lexica and the MT system from parallel data.
documentAligner: DIC
The variable documentAligner
can take three different values, each of them taking a different document-alignment strategy:
DIC
: takes the strategy using bilingual lexica and a linear regressorexternalMT
: takes the strategy using MT, in this case using an external MT script (provided by the user) that reads source-language text from the standard input and writes the translations to the standard outputNMT
: uses parallel data to train a neural MT (NMT) system that is then used for document alignment
dic: /home/user/en-fr.dic
Option dic
allows to specify the path to the bilingual lexicon to be used for document alignment. If the lexicon specified does not exist, the pipeline will try to build it using a parallel corpus provided through the variable initCorpusTrainPrefix
:
initCorpusTrainPrefix: ['/home/user/Europarl.en-fr.train']
This variable must contain one or more corpora prefixes. For a given prefix (/home/user/training
in the example) the pipeline expects to find one file prefix
.lang1
and another prefix
.lang2
(in the example, /home/user/Europarl.en-fr.train.en
and /home/user/Europarl.en-fr.train.fr
). If several training prefixes are provided, the corresponding files will be concatenated before building the bilingual lexicon.
Suggestion: a number of pre-built bilingual lexica is available in the repository bitextor-data. It is also possible to use other lexica already available, such as those in OPUS, as long as their format is the same as those in the repository.
alignerCmd: "example/dummy-translate.sh"
docAlignThreshold: 0.1
alignerCmd
: command to call the external MT scriptdocAlignThreshold
: threshold for discarding document pairs with a very low TF/IDF similarity score; this option takes values in [0,1] and is 0.0 by default
If this option is chosen, a Marian NMT model will be trained and evaluated before using it for document alignment. Note that, given the computational cost of training an NMT system, this option requires having an available GPU. The following are mandatory variables in order to build the NMT system:
initCorpusTrainPrefix: ['/home/user/Europarl.en-fr.train']
initCorpusDevPrefix: ['/home/user/Europarl.en-fr.dev']
initCorpusTestPrefix: ['/home/user/Europarl.en-fr.test']
marianDir: /home/user/marian-dev
mosesDir: /home/user/mosesdecoder
subwordNmtDir: /home/user/subword-nmt
nmtVocabSize: 50000
LANG2Detokenizer: "/home/user/mosesdecoder/scripts/tokenizer/detokenizer.perl -l fr"
gpuId: 0
marianArgs: [" --optimizer-delay 1", "--mini-batch-fit", "--mini-batch 1000", "--maxi-batch 1000", "--overwrite", "--keep-best", "--valid-metrics perplexity", "--valid-log valid.log", "--log train.log", "--dropout-rnn 0.2", "--dropout-src 0.2", "--dropout-trg 0.2 ", "--cost-type ce-mean-words", "--layer-normalization", "--exponential-smoothing", "--tied-embeddings", "--valid-metrics bleu"]
initCorpusTrainPrefix
,initCorpusDevPrefix
, andinitCorpusTestPrefix
: training data prefixes, development data prefixes and test data prefixes. See section Variables for document alignment using bilingual lexica for a description of such prefixesmarianDir
: path to the directory containing the installation of the NMT tool MarianmosesDir
: path to the directory containing the MT tool Moses; note that only data pre-processing scripts are used from Moses and, therefore, it is not necessary to compile the project to use it to train and NMT systemsubwordNmtDir
: path to the directory containing the installation of the tool subword-nmtnmtVocabSize
: size of the NMT vocabularyLANG2Detokenizer
: path to a detokenization script that reads from the standard input and writes to the standard outputgpuId
: id of the GPU to be used for training and testingmarianArgs
: additional arguments for Marian training
After document alignment, the next step in the pipeline is segment alignment. This can be carried out by using the tool hunalign or the tool bleualign. The first one uses a bilingual lexicon and is best suited for the DIC
option of documentAligner
; the second one uses MT and is only available if one of the options based on MT has been specified in documentAligner
.
bleualign: true
bleuAlignThreshold: 0.1
hunalignThreshold: 0.0
bleualign
: if this option is set, bleualign is used instead of hunalign as the tool for segment alignment. This option will only work isdocumentAligner
is set either toexternalMT
orNMT
. This option false by defaultbleuAlignThreshold
andhunalignThreshold
: score threshold for filtering pairs of sentences with a score too low.bleuAlignThreshold
should be set to a value in [0,1], whilehunalignThreshold
can take any float value. Both are set to 0.0 by default
Parallel data filtering is carried out with the tool Bicleaner; this tool uses a pre-trained regression model to filter out pairs of segments with a low confidence score (learn more about Bicleaner at [https://github.com/bitextor/bicleaner]). The options required to make it work are:
bicleaner: /home/user/bicleaner-model/en-fr/training.en-fr.yaml
bicleanerThreshold: 0.6
bicleaner
: path to the YAML configuration file of a pre-trained model. A number of pre-trained models are available at [https://github.com/bitextor/bitextor-data/releases/tag/bicleaner-v1.0]. They are ready to be downloaded and decompressedbicleanerThreshold
: threshold for the confidence score obtained with bitextor to filter low-confidence segment pairs. It is recommended to set it to values in [0.5,0.7], even though it is set to 0.0 by default
If the bicleaner model is not availalbe, the pipeline will try to train one automatically from the data provided through the config file options initCorpusTrainPrefix
and bicleanerCorpusTrainingPrefix
:
initCorpusTrainPrefix: ['/home/user/Europarl.en-fr.train']
bicleanerCorpusTrainingPrefix: '/home/user/RF.en-fr'
initCorpusTrainPrefix
: prefix to parallel corpus (see section Variables for document alignment using bilingual lexica) that will be used to train statistical dictionaries which are part of the bicleaner modelbicleanerCorpusTrainingPrefix
: prefix to the parallel corpus that will be used to train the regressor that obtains the confidence score in Bicleaner It is important to provide different parallel corpora for these two options as this helps bicleaner to deal with unkown words (that do not apear in the statistical dictionaries) during scoring.
Some other options can be configured to specify the output format of our corpus:
elrc: true
tmx: true
deduped: false
elrc
: if this option is set, some ELRC quality indicators are added to the final corpus, sich as length ratio; these indicators can be used later to filter-out some segment pairs manuallytmx
: if this option is set, the output corpus is formatted as a TMX translation memorydeduped
: if this option is set in conjunction withtmx
, the resulting TMX will not contain repeated segment pairs; if a segment pair is found in more than one pair of documents, it will be provided with more than two URLs, so it is possible to know in which original URLs it appeared
Bitextor is a pipeline that runs a collection of scripts to produce a parallel corpus from a collection of multilingual websites. The pipeline is divided in five stages:
- Crawling: documents are downloaded from the specified websites
- Pre-processing: downloaded documents are normalized, boilerplates are removed, plain text is extracted, and language is identified
- Document alignment: parallel documents are identified. Two strategies are implemented for this stage:
- one using bilingual lexica and a collection of features extracted from HTML; a linear regressor combines these resources to produce a score in [0,1], and
- another using machine translation and a TF/IDF strategy to score document pairs
- Segment alignment: each pair of documents is processed to identify parallel segments. Again, two strategies are implemented:
- one using the tool Hunalign, and
- another using Bleualign, that can only be used if the MT-based-document-alignment strategy is used (machine translations are used for both methods)
- Post-processing: final steps that allow to clean the parallel corpus obtained using the tool bicleaner, deduplicates translation units, and computes additional quality metrics
The following diagram shows the structure of the pipeline and the different scripts that are used in each stage: