Proto-word reconstruction with NNs

Get the data

Our self-compiled dataset is included in the repository at data/romance_swadesh*.csv. For the data derived from the Ciobanu 2014 dataset, please email us.

How to train our models

Python requirements

We used Python 3.7.9 in a conda environment. All required packages are listed in requirements.txt. To install them run:

pip install -r requirements.txt

in your local environment.

We expect you to execute the scripts in the source folder.

Command line parameters

This list can be queried in the command line by running a script with the -h flag.

Parameter	Function	Default
--aligned	Use the rows containing the manual alignments	False
--ancestor	Header of the column containing the ancestor of the cognate set	latin
--data	Path to the file containing the cognate sets	../data/romance_swadesh_ipa.csv
--epochs	Number of epochs	10
--model	One of [ipa, asjp, latin]	ipa
--n_hidden	Number of hidden layers in the feedforward model	2
--ortho	Use one-hot character embeddings	False
--out_tag	Flag for the output folder	swadesh

Scripts

All Python scripts can be found in scource/. The scripts ending in _cv.py Perform cross-validation on the dataset with 5 cross-validation folds.

Name	Description	Cross-validation
feedforward.py	Trains the feedforward model	No
many2one_lstm.py	Trains the LSTM model	No
feedforward_cv.py	Trains the feedforward model	Yes
many2one_lstm_cv.py	Trains the LSTM model	Yes
ciobanu_rnn.py	Trains the RNN model	Yes

Sample script calls

• To train the feedforward model on ASJP feature encodings and the aligned data:

python feedforward.py --data=../data/romance_asjp_full.csv --model=asjp --aligned --out_tag=swadesh

The results will be saved at out/plots_swadesh_feedforward.

• To train the LSTM model on Latin character embeddings on dataset A:

python many2one_lstm.py --data../data/romance_ciobanu_latin.csv --model=latin --ortho

The --ortho flag is required here since we don't have feature encodings for the Latin characters.

• To train the feedforward model on IPA character embeddings on dataset B and run cross-validation:

python feedforward_cv.py --data=../data/romance_swadesh_ipa.csv --model=latin --ortho

• To train the RNN model on ASJP feature encodings and the aligned data:

python ciobanu_rnn.py --data=../data/romance_asjp_full.csv --model=asjp --ortho --out_tag=swadesh