sigmorphon2016-system

University of Helsinki contribution to the SIGMORPHON 2016 shared task on morphological reinflection

A description can be found in the system paper from SIGMORPHON (BibTeX).

The system is implemented in Keras, and there are two versions: the shared task submission and the convolutional system which was finished after the shared task deadline, but is described in the paper.

Both programs assume that the shared task data are available in the directory ../sigmorphon2016, which should be cloned from here.

It is recommended to use virtualenv to set up the dependencies of this project. To do so, execute the following commands:

python3 -m venv /path/to/virtual/environment
source /path/to/virtual/environment/bin/activate
pip3 install --upgrade pip
pip3 install -r requirements.txt

Evaluation can be executed with e.g. python3 sigmorphon-conv.py 123 german navajo which will write output to the models directory (which is assumed to exist!) using the experiment identifier 123 and training models for German and Navajo. After you finished testing, execute deactivate to exit the virtual environment.

The particular experiments to run currently have to be hard-coded in run_experiments() -- sorry about this, adding a convenient UI was not a priority for the shared task.