This repo provides a Pytorch implementation of match network and prototypical network for NLP.

Dataset

I make a few-shot dataset from NYT. This dataset contains 989 sentences. The task aims at predicting same entity in sentence.

Acknowledgements

Special thanks to https://github.com/gitabcworld/MatchingNetworks and https://github.com/orobix/Prototypical-Networks-for-Few-shot-Learning-PyTorch for their implementations.

Training

You can train match network and prototypical network with src/mainNYT.py. If you train match network, you should use fit_match_net(); If you train prototypical network, you should use fit_prototypical_net();

$ python mainNYT.py

The script takes the following command line options:

• dataset_root: the root directory where tha dataset is stored, default to '../NYT'

You can ajust other parameters in src/config.py

Performance

                        val acc     test acc
match network           0.4769      0.4423        
prototypical network    0.70625     0.6938