Skip to content


Repository files navigation



Biroamer is a small utility that will help you anonymise or, better said, ROAM (Random, Omit, Anonymize and Mix) your parallel corpus. It will read an input TMX and output a ROAMed TMX. This means that the resulting TMX will have sentences from the input file randomly shuffled and omitted (around of 10% of the setences will be removed), mixed with another corpus, and with named entities highlighted using <hi></hi> tags.

Currently, Biroamer identifies named entities using Flair NER tagger on one side of the corpus (only English has been tested, but other languages could also be used) and tag the equivalent named-entity on the other side of the corpus using word alignments as computed by fast_align.

Installation instructions


git clone --recursive


  • Python >= 3.7
  • GNU Parallel

Build fast_align

Install packages required by fast_align:

sudo apt install libgoogle-perftools-dev libsparsehash-dev

And build it:

mkdir build
cd build
cmake .. -DCMAKE_INSTALL_PREFIX:PATH=/your/prefix/path
make -j all install

Install Python requirements

pip install .

Download Flair model

python -c "from flair.models import SequenceTagger; SequenceTagger.load('flair/ner-english-fast')"

Conda package

If you want to install Biroamer in a conda environment, run:

conda install -c conda-forge -c bitextor -c bioconda biroamer


The script receives a TMX file as an input and outputs another TMX. The needed parameters are lang1 and lang2 (in ISO 639-1 format). Optionally a corpus in Moses format (tab-separated sentences: sent1 \t sent2) in the same language combination used in the parameters can be given to be mixed with the input corpus. Also using -o will randomly omit about 10% of the sentences from the input corpus.

Usage: biroamer [options] <lang1> <lang2>
    -s SEED           Set random seed for reprodibility
    -a ALIGN_CORPUS   Extra corpus to improve alignment
                      It won't be included in the output
    -j JOBS           Number of jobs to run in parallel
    -b BLOCKSIZE      Number of lines for each job to be processed
    -m MIX_CORPUS     A corpus to mix with
    -o                Enable random omitting of sentences
    -t TOKL1          External tokenizer command for lang1
    -T TOKL2          External tokenizer command for lang2
    -h                Shows this message

If the input corpus plus the mixing corpus are not big enough (at least 100K sentences) to compute word alignments to tag named entities in the other side of the corpus, it is advised to use the -a option to add more sentences and improve this alignment.

If your mixing corpus is in TMX format, you can use tmxt (included in this repository) to obtain a sample of size $SIZE in the aforementioned Moses format:

$ cat mixing-corpus.tmx | python tmxt/ --codelist l1,l2 | head -$SIZE > mix-corpus.txt


With en-es-file.tmx being an input TMX file containing translation units like:

    <tuv xml:lang="en">
        <seg>The e-mail address of John Doe is</seg>
    <tuv xml:lang="es">
        <seg>El correo electrónico de John Doe es</seg>

Mixing corpus mix-corpus-en-es.txt being:

Can you trust your neighbours?        ¿Puedes confiar en tus vecinos?
Bert and Margaret raised seven sons in the 50's.       Bert y Margaret criaron siete hijos en los 50.

And after running the following command:

$ cat en-es-file.tmx | biroamer -o -m mix-corpus-en-es.txt en es > result-en-es.tmx

Will result in results-en-es.tmx being like:

    <tuv xml:lang="en">
        <seg><hi>Bert</hi> and <hi>Margaret</hi> raised seven sons in the 50's.</seg>
    <tuv xml:lang="es">
        <seg><hi>Bert</hi> y <hi>Margaret</hi> criaron siete hijos en los 50.</seg>
    <tuv xml:lang="en">
        <seg>The e-mail address of <hi>John Doe</hi> is <hi></hi></seg>
    <tuv xml:lang="es">
        <seg>El correo electrónico de <hi>John Doe</hi> es <hi></hi></seg>
    <tuv xml:lang="en">
        <seg>Can you trust your neighbours?</seg>
    <tuv xml:lang="es">
        <seg>¿Puedes confiar en tus vecinos?</seg>

External tokenizer command can be used with -t and -T for lang1 and lang2 respectively. For example:

$ cat en-es-file.tmx \
    | biroamer \
        -o -m mix-corpus-en-es.txt \
        -t "mosesdecoder/scripts/tokenizer/tokenizer.perl -l en -no-escape" \
        -T "mosesdecoder/scripts/tokenizer/tokenizer.perl -l es -no-escape" \
        en es \
        > result-en-es.tmx

But it is recommended to use the default tokenizer unless you are working with a language that NLTK does not support. Also note that the tokenizer command is already parallelized inside biroamer using parallel, so it is advised to use single-threaded commands.

Disable GPU

By default, Flair uses GPUs if available. If you want to disable GPU processing or choose which GPUs should use, you just need to set CUDA_VISIBLE_DEVICES environment variable.

To avoid using GPU:

$ cat en-es-file.tmx | CUDA_VISIBLE_DEVICES="" biroamer -o -m mix-corpus-en-es.txt en es > result-en-es.tmx

or use only 2 of 4 available:

$ cat en-es-file.tmx | CUDA_VISIBLE_DEVICES="0,1" biroamer -o -m mix-corpus-en-es.txt en es > result-en-es.tmx

Connecting Europe Facility

All documents and software contained in this repository reflect only the authors' view. The Innovation and Networks Executive Agency of the European Union is not responsible for any use that may be made of the information it contains.