Repo for the Transducer model we used in the paper Morphological Inflection with Phonological Features. This is actually a fork of the model by Makarov & Clematide, adjusted to the 2 new formats that use phonology. Also, it is converted to Python3.
See the dependencies at requirements.txt.
For Linux, run dummy.sh. Configure the 4 arguments and the paths at defaults.py.
For Windows, run run_transducer.py with the relevant arguments.
The script expects train, dev & test files to be located at DATA_PATH
If you wish to experiment with a new language, provide a list of the alphabet and their corresponding phonemes. Optionally, you can add methods for manual conversions (digraphs and trigraphs are supported). Also, you could provide a method for data normalizing.
To reformat words as sequences of phonological features, use the flag --use-phonology. To model them as self-representation of phonemes, use also the flag --self-attn.