robust-maml

Code to reproduce the results described in the paper "Minimax and Neyman-Pearson Meta-Learning for Outlier Languages" (aclweb).

1. Create conda env: conda create -n pos-bert python=3.7
2. Install dependencies: pip install -r requirements.txt
3. Run data creation scripts (make_pos_data.py and make_qa_data.py. See top comments in files for more details)
4. Use the configs from the configs folder or suitably modify them as required to start training

To start training, simply pass the config path to trainer.py

$ python trainer.py --config_path=<config-path>

To run fast adapt, and evaluate the trained models on one outlier-language, run

$ python tester.py --test_lang=<lang> --model_path=<saved-model-path>

To run the tester on all outlier-languages, use run_test.sh. It contains two functions, run_pos and run_qa which will evaluate POS tagging and QA on all languages, respectively. The individual language results will be stored in <saved-model-path>/result. To merge them all into a single CSV file, run

$ python create_csv.py --model_path=<saved-model-path>

For QA, since we split the languages into two groups and train two models. Therefore, create_csv.py expects 2 paths: --model_path and --model_path2

Citation

@inproceedings{ponti-etal-2021-minimax,
    title = "Minimax and Neyman{--}{P}earson Meta-Learning for Outlier Languages",
    author = "Ponti, Edoardo Maria  and
      Aralikatte, Rahul  and
      Shrivastava, Disha  and
      Reddy, Siva  and
      S{\o}gaard, Anders",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-acl.106",
    pages = "1245--1260",
}