Transformed Protoform Reconstruction

Installation

pip install -r requirements.txt Then separately install torch (https://pytorch.org/get-started/locally/)

Phylogeny probe dependencies

To run the phylogeny probe, you may need to manually install tqDist and PHYLIP.

• tqDist - used to calculate generalized quartet distance (GQD)

On a Linux machine, conda install -c bioconda tqdist or Follow the instructions on https://www.birc.au.dk/~cstorm/software/tqdist/

• PHYLIP consense - used to get a consensus tree Install PHYLIP from https://evolution.genetics.washington.edu/phylip/ (use wget if needed) Follow instructions at https://evolution.genetics.washington.edu/phylip/install.html Copy ./consense to src/phylogeny/: cp PATH_TO_PHYLIP/phylip-3.695/exe/consense.app/Contents/MacOS/consense src/phylogeny

Running our code

Note: Do not run the code on Mac M1 silicon - the results will not hold.

Training

export WORK_DIR=.           
export SRC_DIR=src
export DATA_DIR=data
export CONF_DIR=conf

python src/train.py --conf chinese_baxter.json
python src/train.py --conf romance_orto_meloni.json
python src/train.py --conf romance_ipa_meloni.json

Evaluation

export WORK_DIR=.           
export SRC_DIR=src
export DATA_DIR=data
export CONF_DIR=conf

python src/evaluate.py --conf romance_ipa_meloni
python src/evaluate.py --conf romance_orto_meloni
python src/evaluate.py --conf chinese_baxter

If you would like to run the evaluation with romance_ipa and romance_orto (the full datasets we use in our main results), please contact Ciobanu and Dinu (2014) for their data.

Phylogeny probe:

python src/phylogeny_probe.py --model transformer --dataset romance_orto
python src/phylogeny_probe.py --model transformer --dataset romance_ipa
python src/phylogeny_probe.py --model transformer --dataset chinese_baxter
. ./src/phylogeny/phylogeny.sh

(don't forget the dot)

RNN baseline (re-implementation of Meloni et al)

This PyTorch re-implementation of Meloni et al 2021 supersedes our previous re-implementation at https://github.com/cmu-llab/meloni-2021-reimplementation.

Training:

wandb login

export WORK_DIR=.
export DATA_DIR=data

python rnn/train_encoder_decoder_rnn.py \
    --dataset DATASET --save_predictions true \
    --lr \
    --beta1 \
    --beta2 \
    --encoder_layers 1 \
    --decoder_layers 1 \
    --embedding_size \
    --dim_feedforward \
    --dropout \
    --epochs \
    --lang_separators \
    --warmup_epochs \
    --batch_size \
    --wandb_name YOUR_WANDB_PROJECT \
    --wandb_entity YOUR_WANDB_ORG \
    --sweeping true

(note - do not set the batch size to anything besides 1 for decoding)

Evaluation:

python rnn/evaluate.py --model rnn --dataset romance_orto_meloni
python rnn/evaluate.py --model rnn --dataset romance_ipa_meloni
python rnn/evaluate.py --model rnn --dataset chinese_baxter

If you would like to run the evaluation with romance_ipa and romance_orto (the full datasets we use in our main results), please contact Ciobanu and Dinu (2014) for their data.

Phylogeny probe:

python rnn/phylogeny_probe.py --model rnn --dataset romance_orto
python rnn/phylogeny_probe.py --model rnn --dataset romance_ipa
python rnn/phylogeny_probe.py --model rnn --dataset chinese_baxter

Citing our paper

Please cite our paper as follows:

Young Min Kim, Kalvin Chang, Chenxuan Cui, and David Mortensen (forthcoming). Transformed Protoform Reconstruction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL 2023), Toronto, Canada.

@InProceedings{Kim-et-al:2023,
  author = {Kim, Young Min and Chang, Kalvin and Cui, Chenxuan and Mortensen, David R.},
  title = {Transformed Protoform Reconstruction},
  booktitle = {Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL 2023)},
  year = {forthcoming},
  month = {July},
  date = {9--14},
  location = {Toronto, Canada},
}