GAlign is yet another word alignment tool (like GIZA++, MGiza, fast_align,...) with several distinguishing features:

• competitive alignment quality
• supports incremental training and forced alignment
• fast performance thanks to multi-threading in all stages, scales to many cores
• easy to use
• simple code base with minimal dependencies (Boost, Intel TBB)

Download and compile GAlign:

git clone https://github.com/ales-t/galign.git
cd galign
# export BOOST_PATH=/optionally/specify/different/path/to/Boost
make

Verify that you can run the program:

bin/wordalign --help

Use of GAlign can be roughly divided into 3 modes:

Basic usage:

paste corpus.src corpus.tgt | bin/wordalign -m > corpus.alignment

For greater speed, you might like to skip the final Viterbi search for the best alignment (and use the last Gibbs sample instead) with the switch --no-viterbi.

Train and store the model:

paste corpus.src corpus.tgt | bin/wordalign -m --store-model-file corpus.model > corpus.alignment

In this mode, the existing model will be loaded and used as a starting point for training. Note that data size matters: if the old data are very small compared to the new corpus, statistics collected from the new data will outweigh the existing model, and vice versa.

paste newcorpus.src newcorpus.tgt | bin/wordalign -m --load-model-file oldcorpus.model > newcorpus.alignment

Do not train, only search for the best alignment given the existing model. Avoid --no-viterbi in this mode.

paste newcorpus.src newcorpus.tgt | bin/wordalign -f -m --load-model-file oldcorpus.model > newcorpus.alignment

Author: Aleš Tamchyna <tamchyna -at- ufal.mff.cuni.cz>

Copyright © 2013 Institute of Formal and Applied Linguistics, Charles University in Prague.

Licensed under the GNU Lesser General Public License version 2.1.