phonotactic-complexity

This repository contains code for analysing phonotactics. It accompanies the paper: Phonotactic Complexity and Its Trade-offs (Pimentel et al., TACL 2020).

It is a study about languages phonotactics and how it relates to other language features, such as word length.

Install Dependencies

Create a conda environment with

$ source config/conda.sh

And install your appropriate version of PyTorch.

Parse data

First, download NorthEuraLex data from this link and put it in the datasets/northeuralex folder. Then, parse it using the following command:

$ python data_layer/parse.py --data northeuralex

Train models

Train base models

You can train the base models (without shared embeddings) with the commands:

$ python learn_layer/train_base.py --model <model> [--opt]
$ python learn_layer/train_base_bayes.py --model <model>
$ python learn_layer/train_base_cv.py --model <model> [--opt]

Different commands are:

• train_base: Trains model with default data split;
• train_base_bayes: Trains model using bayesing optimization and default data split;
• train_base_cv: Trains cross validated models.

Model can be:

• lstm: LSTM with default one hot embeddings
• phoible: LSTM with phoible embeddings
• phoible-lookup: LSTM with both one hot and phoible embeddings

And --opt is an optional parameter that tells the script to use bayes optimized hyper-parameters. It can only be used after training model with train_base_bayes.

Train shared models

You can train models with shared embeddings using the commands:

$ python learn_layer/train_shared.py --model <model> [--opt]
$ python learn_layer/train_shared_bayes.py --model <model>
$ python learn_layer/train_shared_cv.py --model <model> [--opt]

Model can be:

• shared-lstm: LSTM with shared one hot embeddings
• shared-phoible: LSTM with shared phoible embeddings
• shared-phoible-lookup: LSTM with both one hot and phoible shared embeddings

Train ngram models

You can train ngram models with the following commands:

$ python learn_layer/train_ngram.py --model ngram
$ python learn_layer/train_unigram.py --model unigram

$ python learn_layer/train_ngram_cv.py --model ngram
$ python learn_layer/train_unigram_cv.py --model ngram

Model can be:

• ngram: ngram model by default is a trigram
• unigram: Unigram model

Train models on artificial data

You can train models on aritificial data using the commands:

$ python learn_layer/train_artificial.py --artificial-type <artificial-type>
$ python learn_layer/train_artificial_bayes.py --artificial-type <artificial-type>
$ python learn_layer/train_artificial_cv.py --artificial-type <artificial-type>
$ python learn_layer/train_artificial_ngram.py --model ngram --artificial-type <artificial-type>

Artificial type can be:

• harmony: Artificial data with vowel harmony removed;
• devoicing: Artificial data with final obstruent devoicing removed.

Train all models

You can also call a script to train all models sequentially (it might take a while):

$ source learn_layer/train_multi.sh

Plot Results

Get compiled result data:

$ python analysis_layer/compile_results.py
$ python analysis_layer/get_lang_inventory.py

Plot all results with commands:

$ mkdir plot
$ python visualization_layer/plot_lstm.py
$ python visualization_layer/plot_full.py
$ python visualization_layer/plot_inventory.py
$ python visualization_layer/plot_kde.py
$ python visualization_layer/plot_artificial_scatter.py

Extra Information

Citation

If this code or the paper were usefull to you, consider citing it:

@article{pimentel-etal-2020-phonotactics,
    title={Phonotactic Complexity and its Trade-offs},
    author={Pimentel, Tiago and Roark, Brian and Cotterell, Ryan},
    journal={Transactions of the Association for Computational Linguistics},
    volume={8},
    pages={1--18},
    year={2020},
    publisher={MIT Press},
    doi={10.1162/tacl\_a\_00296},
    url={https://www.mitpressjournals.org/doi/abs/10.1162/tacl_a_00296}
}

Dependencies

This project was tested with libraries:

numpy==1.15.4
pandas==0.24.1
scikit-learn==0.20.2
tqdm==4.31.1
matplotlib==2.0.2
seaborn==0.9.0
torch==1.0.1.post2

Contact

To ask questions or report problems, please open an issue.