© [2024] Ubisoft Entertainment. All Rights Reserved

BinaryAlign: Word Alignment as Binary Sequence Labeling

Introduction

BinaryAlign reformulates word alignment as a set of binary sequence labeling tasks. It outperforms existing approaches in both high and low-resource language settings, providing a unified approach to word alignment. This repository contains the code and models for BinaryAlign as described in our paper accepted to the main conference of ACL 2024.

Datasets

Format

• source (.src)

He has a sofa .

• target (.tgt)

Il a un canapé .

• gold alignment (.talp)

1-1 2-2 3-3 4-4 5-5 

Data

We used the same datasets as https://github.com/sufenlp/AccAlign.

Training

bash train.sh

Evaluation

bash eval.sh

Checkpoints

Training Languages	Link
Align6	models/align6
deen	models/deen
roen	models/roen
fren	models/fren
zhen	models/zhen
jaen	models/jaen

Citation

@article{latouche2024binaryalign,
  title={BinaryAlign: Word Alignment as Binary Sequence Labeling},
  author={Latouche, Gaetan Lopez and Carbonneau, Marc-Andr{\'e} and Swanson, Ben},
  journal={arXiv preprint arXiv:2407.12881},
  year={2024}
}

License

See Licence File - CC4.0 non commercial

© [2024] Ubisoft Entertainment. All Rights Reserved