README.md

Official Code for "ZGUL: Zero-shot Generalization to Unseen Languages using Multi-source Ensembling of Language Adapters"

1. Environment Setup

TBD

2. Zero-shot Inference

EM Steps (T) and LR (lr) for each target language (tuned on closest language dev set) along with Test F1 scores can be found in em_params_zero_shot.png . Please run as follows:

First copy infer* files from `scripts' directory to current one

• Germanic

bash infer_germanic.sh

• Slavic

bash infer_slavic.sh

• African

bash infer_african.sh

• Indic

bash infer_indic.sh

3. Training Instructions

First copy train* files from `scripts' directory to current one

• Germanic

bash train_udpos.sh en,is,de en,is,de

• Slavic

bash train_udpos.sh en,ru,cs en,ru,cs

• African

bash train_masa.sh en,amh,swa,wol en_conll,am,sw,wo

• Indic

bash train_panx.sh en,hi,bn,ur en_ner,hi,bn,ur

4. Trained model checkpoint

link

Cite

The codebase is a part of the work ZGUL: Zero-shot Generalization to Unseen Languages using Multi-source Ensembling of Language Adapters. If you use or extend our work, please cite the following paper:

@inproceedings{rathore-etal-2023-zgul,
    title = "{ZGUL}: Zero-shot Generalization to Unseen Languages using Multi-source Ensembling of Language Adapters",
    author = "Rathore, Vipul  and
      Dhingra, Rajdeep  and
      Singla, Parag  and
      {Mausam}",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.431",
    pages = "6969--6987",
}

Acknowledgements

Our codebase is built upon Adapterhub's. For more details on the transformers source code used, we refer the user to their repository.

For more details on the dataset and training scripts used, we refer the user to Google xtreme repo.