apertium-lexical-training

This is a script for training lexical selection rules based on parallel or non-parallel corpora, for use in Apertium machine translation pairs.

The procedure for lexical selection training has historically been a bit messy, with various scripts involved that require lots of manual tweaking, and many third party tools to be installed. The goal of this project is to make the training procedure as streamlined and user-friendly as possible.

Requirements

Training on parallel corpora:

If you have a parallel corpus, you should use that.

• parallel corpus
• apertium-core (install apertium-lex-tools with yasmet)
• fast_align
• language pair (install locally)
• python dependencies in requirements.txt

Training on non-parallel corpora:

If you have don't have a parallel corpus, you can still get useful rules out of a monolingual corpus.

• non-parallel corpus
• apertium-core
• language pair (install locally)
• IRSTLM
• python dependencies in requirements.txt

Optional dependencies:

• pv – gives you a progress bar for some of the slower steps

Installing requirements on Ubuntu

We assume you've already got Tino's nightly repo set up, since you're developing a language pair – if not, see http://wiki.apertium.org/wiki/Ubuntu.

sudo apt install -y irstlm pv

The python dependencies aren't currently packaged for Ubuntu etc., so to install them, you should use a VirtualEnv to not clutter your regular Python path:

sudo apt install python3-virtualenv
virtualenv .venv
source .venv/bin/activate  # you have to do this every time you want the dependencies to be available
.venv/bin/pip3 install -r requirements.txt

For parallel training, you'll also need fast_align; following config.toml.example we'll check it out in the parent directory (so it's "next to" this project):

sudo apt install -y libgoogle-perftools-dev libsparsehash-dev

git clone https://github.com/clab/fast_align
mkdir -p fast_align/build
cd fast_align/build
cmake ..
make -j

Finding corpora

Parallel corpora

If you don't have a corpus, the first place you should go looking is Opus. Pick a source and target language, then download the file moses column if there is one. The moses column gives you a parallel, sentence-aligned corpus. You'll get a zip file containing a pair of files with aligned lines, these are the files you use in training.

Say you downloaded the WikiMatrix Moses file for Norwegian and Swedish, you'll get a zip file called no-svn.txt.zip. Unzip that, and you'll see files

LICENSE
README
WikiMatrix.no-sv.no
WikiMatrix.no-sv.sv
WikiMatrix.no-sv.xml

The files ending in the language codes contain the actual sentences. We can verify that they're aligned sentences with

paste WikiMatrix.no-sv.no WikiMatrix.no-sv.sv |less -S

If you found nothing on Opus, you may want to check Corpora on the Apertium wiki, or try using a non-parallel corpus.

Non-parallel corpora

Any text in a .txt file will do, as long as it's in your target language. See Corpora on the Apertium wiki for possible sources.

You can also use a precompiled (target) language model if you have that, instead of a corpus.

How to train

• Install the requirements
• Clone this repo
  

  git clone https://github.com/apertium/apertium-lexical-training.git
  

• Create config.toml (for reference, see config.toml.example)

  cp config.toml.example config.toml
  

  • Now edit config.toml and give the rigth paths to tools and your corpus
  • Enter IS_PARALLEL = true for parallel corpora, false for non-parallel corpora
• Run lexical_training.py
  

  python3 lexical_training.py [CONFIG_FILE]
  
  args:
    CONFIG_FILE : optional, default='config.toml'
  

Tests

This folder contains scripts and data for automated testing of the training scripts. It is run as a Github workflow with the little test corpus included in the repo.

References

This project was started as part of Google Summer of Code 2021 (complete work submission: https://apertium.projectjj.com/gsoc2021/vivekvardhanadepu/vivekvardhanadepu.html , original GsoC Idea https://wiki.apertium.org/wiki/Ideas_for_Google_Summer_of_Code/User-friendly_lexical_selection_training ).

[1] Philipp Koehn. Europarl: A Parallel Corpus for Statistical Machine Translation. MT Summit 2005.

[2] https://www-i6.informatik.rwth-aachen.de/web/Software/YASMET.html