This is the repository for the paper "Breaking the Script Barrier in Multilingual Pre-Trained Language Models with Transliteration-Based Post-Training Alignment", which aims to improve the cross-lingual transfer of multilingual models by better aligning the representations of languages written in different scripts. We utilize a rule-based transliteration system (uroman) to generate Latin-script transliterations of sentences in different scripts, and then use our proposed post-pretraining alignment (PPA) method to fine-tune a pretrained multilingual model on the paired data.
Our method aims to better align cross-script representations at sentence and token levels by leveraging transliteration to a common script. We use Glot500, a mPLM pretrained on over 500 languages, as our source model, and fine-tune it on datasets formed from two different language groups ( Mediterranean-Amharic-Farsi and South+East Asian Languages ). The languages in each group share areal features but are written in different scripts. We evaluate the fine-tuned models on a series of downstream tasks, and show that after PPA, the models consistently outperform the original Glot500 model when trasnferring from different source languages.
Paper on arXiv: https://arxiv.org/abs/2406.19759
- Mediterranean-Amharic-Farsi: https://huggingface.co/orxhelili/translit_ppa_mediterranean
- South+East Asian Languages: https://huggingface.co/orxhelili/translit_ppa_se_asian
Given a text file containing sentences in different language-scripts, use the following command to generate a csv file where each row is a pair of sentences (in its original script and in the Latin script).
python utils/preprocess_dataset.py
An example script for fine-tuning the model based on SLURM is provided in sbatch_scripts/sbatch_finetune.sh.
Please refer to Glot500 and SIB200 for downloading the datasets used for evaluation.
Example scripts based on SLURM for running the different evaluations are provided in sbatch_scripts/.
If you find our models useful for your research, please considering citing:
@article{xhelili2024breaking,
title={Breaking the Script Barrier in Multilingual Pre-Trained Language Models with Transliteration-Based Post-Training Alignment},
author={Xhelili, Orgest and Liu, Yihong and Sch{\"u}tze, Hinrich},
journal={arXiv preprint arXiv:2406.19759},
year={2024}
}