This repository contains code accompanying 'Equivariant Transduction through Invariant Alignment' by Jennifer C. White and Ryan Cotterell (COLING 2022).
Run pip install -r requirements.txt to install all requirements.
Download SCAN and place it in the directory.
Running python train.py --best_hyperparams SPLIT_NAME will train a model on the given split using our best-performing hyperparameters for that split.
SCAN Splits available are simple, add_jump, length_generalization and around_right.
Custom hyperparameters can also be passed in as arguments. Run python train.py -h to see information on these arguments.
Following training, the model with lowest loss on the dev set will be evaluated on the test set and saved with a name in the format {test_accuracy}___{scan_split}__{K}__{num_filters}__{hidden_size}__{embed_dim}__{batch_size}__{learning_rate}.model.
Running python test.py --model_path /path/to/model --scan_split SPLIT_NAME will evaluate a model on the test set of the given split.
@inproceedings{white-cotterell-2022-equivariant,
title = "Equivariant Transduction Through Invariant Alignment",
author = "White, Jennifer C. and
Cotterell, Ryan",
booktitle = "Proceedings of the 29th International Conference on
Computational Linguistics",
month = oct,
year = "2022",
publisher = "International Committee on Computational Linguistics",
}