This is the repository for my 2021 paper "Restoring The Sister: Reconstructing a Lexicon from Sister Languages using Neural Machine Translation". The actual paper can be found here.
The repository contains the initial data file and all data-processing scripts needed to replicate the experiment. If you are just interested in the data file, please refer to "romance-ortography.txt".
This experiment requires python 3 and the following python dependencies:
- opennmt-py 2.0.1 (other versions should work, but no guarantees can be made)
- jellyfish 0.8.2 (other versions should work)
If you do not have these packages installed, call "pip install --user opennmt-py==2.0.1" followed by "pip install --user jellyfish==0.8.2".
If you wish to replicate the experiment, clone the repository and run the following scripts in this order (please note, these scripts were written tailored to this data, so they will not work for your custom data):
- mungenormalizev2.py (seperates the file into multiple files for each language, extracts vocabulary)
- restructure.py (removes all cognate pairs that do not have an equivalent in Italian, which is our target)
- concatenate.py (concatenates the training data in the manner described in the paper)
- reduce.py (artifically reduces a set of training data for comparison)
After running these scripts, you can call "python execute_all.py --dir here". This will train models for all .yaml files in the directory (warning: this may take up to 12 hours or more). If you want to train individual models, simply call "onmt_train --config config-file".
The trained models will save in the folder /models/. In order to evaluate a model, you need to call "onmt_translate -model model directory and name -src test file for that model -output where you want to save the output". All data, including the test-files is stored in /data/. If you are unsure what the right test-file is for a model, then enter its .yaml file. The test-file will have the same name as the validation target file, except the "_val" will be changed to "_test".
In order to evaluate a model, you can use the "evaluation_command.py" script. It takes two arguments, your translated file (--src) and the target test file (--tgt). Your target file will always be "data/italian_test_res.txt". This script will save a text file which will be named after your translated file + "edit_eval.txt". It will contain edit distances of 0-5 in absolute numbers, the custom score (see paper), and the edit distances in percentages.