Few-Shot Learning with Siamese Networks and Label Tuning

A few-shot learning method based on siamese networks.

Code & models for the paper to appear at ACL 2022.

The Symanto Few-Shot Benchmark

symanto-fsb implements the benchmark discussed in the paper.

It can be easily extended to evaluate new models. See the extension section below.

Installation

pip install -e .

CLI

Char-SVM

symanto-fsb evaluate-char-svm output/char_svm --n-trials=1

This will run the specified number of trials on each dataset and write results to the output directory. Afterwards you can create a result table:

symanto-fsb report output /tmp/report.tsv

Sentence-Transformer

Zero-Shot

symanto-fsb \
   evaluate-sentence-transformer \
    output/pml-mpnet \
    --gpu 0 \
    --n-examples=0 \
    --n-trials=1 

Few-Shot

symanto-fsb \
   evaluate-sentence-transformer \
    output/pml-mpnet \
    --gpu 0 \
    --n-examples=8 \
    --n-trials=1

Extension

In general a new model is added by adding:

1. A new implementation of the Predictor interface
2. A new command to cli.py

Known Issues

Datasets hosted on Google Drive do not work right now: datasets issue/3809

Testing & Maintenance

pip install -r dev-requirements.txt
dev-tools/format.sh
dev-tools/lint.sh
dev-tools/test.sh

Models

For the sake of comparability we trained 4. The code above can be used with the Siamese network models.

Cross Attention

• mpnet-base-snli-mnli (English)
• xlm-roberta-base-snli-mnli-anli-xnli (ML)

Siamese Networks

• sn-mpnet-base-snli-mnli (English)
• sn-xlm-roberta-base-snli-mnli-anli-xnli (ML)

Disclaimer

This is not an official Symanto product!

How to Cite

@inproceedings{labeltuning2022,
	title        = {{Few-Shot} {Learning} with {Siamese} {Networks} and {Label} {Tuning}},
	author       = {M{\"u}ller, Thomas and Pérez-Torró, Guillermo and Franco-Salvador, Marc},
	year         = {2022},
	booktitle    = {ACL (to appear)},
	url          = {https://arxiv.org/abs/2203.14655},
}