Meta-Learning for NER

This is built upon the base-code for the paper Learning to Learn to Disambiguate: Meta-Learning for Few-Shot Word Sense Disambiguation. The code will be updated soon to contain more experiments.

Getting started

• Clone the repository: git clone git@github.com:Nithin-Holla/MetaWSD.git.
• Create a virtual environment.
• Install the required packages: pip install -r MetaWSD/requirements.txt.
• Create a directory for storing the data: mkdir data.
• Navigate to the data directory: cd data.
• Copy the ontonotes-bert directory into the data folder.
• Navigate back: cd ..

Preparing the data

• Make sure train.txt, val.txt and test.txt are in the ontonotes-bert folder
• The labels-train.txt and labels-test.txt indicate the entity classes for training episodes and test episodes respectively.

Training the models

• The YAML configuration files for all the models are in config/wsd. To train a model, run python MetaWSD/train_ner.py --config CONFIG_FILE.
• Training on multiple GPUs is supported for the MAML variants only. In order to use multiple GPUs, specify the flag --multi_gpu.

Troubleshooting

(Already done. No need to do it again.)

If you have a RuntimeError with Proto(FO)MAML and BERT, you can install the higher library from this fork: https://github.com/Nithin-Holla/higher, which has a temporary fix for this. Also, replace diffopt.step(loss) with diffopt.step(loss, retain_graph=True) in models/seq_meta.py.

Citation

If you use this code repository, please consider citing original paper that implemented the base code of our project:

@article{holla2020metawsd,
  title={Learning to Learn to Disambiguate: Meta-Learning for Few-Shot Word Sense Disambiguation.},
  author={Holla, Nithin and Mishra, Pushkar and Yannakoudakis, Helen and Shutova, Ekaterina},
  journal={arXiv preprint arXiv:2004.14355},
  year={2020}
}