This is the home of my bachelor thesis, IPA alignment using vector representations, which presents several different methods for obtaining embeddings of IPA segments in vector space for the purposes of aligning IPA-encoded sound sequences. The repo contains the code implementation, the training and evaluation data, and the thesis paper.
# clone this repository
git clone https://github.com/pavelsof/ipavec
cd ipavec
# create a python3 virtual environment
python3 -m venv meta/venv
source meta/venv/bin/activate
# install the dependencies
pip install -r requirements.txt
# check the unit tests
python -m unittest discover code# activate the virtual env if it is not already
source meta/venv/bin/activate
# ensure reproducibility of the results
export PYTHONHASHSEED=42
# use train.py to produce embeddings for one of the trainable methods
python train.py --help
# use run.py to run a method on a dataset and output aligned pairs
python run.py --help
# use eval.py to evaluate the output if you have the gold-standard alignments
python eval.py --helppython run.py data/bdpa/slavic.psa --vectors phoible --output output/slavic-phoible.psa
python eval.py data/bdpa/slavic.psa output/slavic-phoible.psa | lesspython train.py phon2vec data/northeuralex/ipa --output models/phon2vec
python run.py data/bdpa/slavic.psa --vectors phon2vec --output output/slavic-phon2vec.psa
python eval.py data/bdpa/slavic.psa output/slavic-phon2vec.psa | lesspython train.py nn data/northeuralex/ipa --extra epochs=10,batch_size=128
python run.py data/bdpa/slavic.psa --vectors nn --output output/slavic-nn.psa
python eval.py data/bdpa/slavic.psa output/slavic-nn.psa | less
python train.py rnn data/northeuralex/word_pairs --extra batch_size=128,from_model=models/nn
python run.py data/bdpa/slavic.psa --vectors rnn --output output/slavic-rnn.psa
python eval.py data/bdpa/slavic.psa output/slavic-rnn.psa | lessThe table summarises the accuracy achieved by the implemented methods on the BDPA datasets. The last two columns list the results of the PMI and SCA methods.
| one-hot | phoible | phon2vec | nn | nn+rnn | pmi | sca | |
|---|---|---|---|---|---|---|---|
| covington | 60.61% | 82.42% | 80.18% | 82.52% | 82.52% | 87.80% | 90.24% |
| andean | 85.66% | 87.31% | 97.25% | 99.34% | 99.50% | 95.21% | 99.67% |
| bai | 52.55% | 62.77% | 61.25% | 74.72% | 75.52% | - | 83.45% |
| bulgarian | 60.54% | 80.54% | 77.98% | 82.55% | 86.70% | 81.70% | 89.34% |
| dutch | 14.16% | 25.65% | 26.00% | 32.50% | 32.50% | 36.67% | 42.20% |
| french | 42.94% | 62.92% | 68.94% | 74.30% | 77.04% | 71.98% | 80.90% |
| germanic | 39.93% | 51.78% | 54.59% | 71.83% | 72.55% | 75.32% | 83.48% |
| japanese | 53.56% | 65.04% | 73.74% | 62.71% | 71.08% | 68.26% | 82.19% |
| norwegian | 59.39% | 78.87% | 73.69% | 83.53% | 89.06% | 78.11% | 91.77% |
| ob-ugrian | 59.58% | 77.87% | 73.35% | 78.04% | 82.55% | 82.09% | 86.04% |
| romance | 40.48% | 71.28% | 63.16% | 76.37% | 77.55% | 84.51% | 95.62% |
| sinitic | 27.34% | 28.57% | 30.75% | 72.46% | 74.04% | - | 98.95% |
| slavic | 76.96% | 90.73% | 84.22% | 89.89% | 96.81% | 89.36% | 94.15% |
| global | 51.83% | 66.64% | 66.99% | 75.88% | 78.45% | - | 84.84% |
Copyright (C) 2018 Pavel Sofroniev and Çağri Çöltekin
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see https://www.gnu.org/licenses/.