HMM-Aligner

This is the implementation of word aligner using Hidden Markov Model.

Cite this if you used it:

@article{Gū2018a,
    title = {An Easily Extensible Word Aligner},
    author = {Jetic Gū and Anahita Mansouri-Bigvand and Anoop Sarkar},
    url = {https://ufal.mff.cuni.cz/pbml/111/art-gu-et-al.pdf},
    doi = {10.2478/pralin-2018-0008},
    issn = {0032-6585},
    year = {2018},
    date = {2018-10-01},
    journal = {The Prague Bulletin of Mathematical Linguistics},
    volume = {111},
    pages = {87--96},
    pubstate = {published},
    tppubtype = {article}
}

Note

The HMM Aligner is still under development, so new features and optimisations will be added. However we always maintain a master branch that works. To see how it works please checkout our Wiki page. For development progress please checkout the Project page.

Currently the master branch uses Numpy to speed up the training. They are however unfortunately a bit harder to read of course, so the old versions are kept in src/models/old. If you copy these models out they will be able to run as they are using the same API as the Numpy versions, there are minor differences in the decoding part but the training code are essentially the same, if you want to understand how our aligner works they should be very useful.

Get Started

To use the models starting with c, for example cIBM1, one must compile the files needed.

> cd src
> python setup.py build_ext --inplace

The other models directly under the directory src/models can be directly used.

Usage

> cd src
> python aligner.py -h

For detailed specifications, please checkout our Wiki page for API specs.