Nearest Neighbour Few-Shot Learning for Cross-lingual Classification

Even though large pre-trained multilingual models (e.g. mBERT, XLM-R) have led to significant performance gains on a wide range of cross-lingual NLP tasks, success on many downstream tasks still relies on the availability of sufficient annotated data. Traditional fine-tuning of pre-trained models using only a few target samples can cause over-fitting. This can be quite limiting as most languages in the world are under-resourced. In this work, we investigate cross-lingual adaptation using a simple nearest neighbor few-shot (<15 samples) inference technique for classification tasks. We experiment using a total of 16 distinct languages across two NLP tasks- XNLI and PAWS-X. Our approach consistently improves traditional fine-tuning using only a handful of labeled samples in target locales. We also demonstrate its generalization capability across tasks.

Implementation

This repository contains the code for Nearest Neighbour Few-Shot Learning for Cross-lingual Classification.

Implementation of the Nearest Neighbor Few-Shot Learning approach and instructions on running the code will be available here, soon.

Environment Creation Commands

We provide a *.yml file of our environment. Install environment by,

conda env create -f scripts/few-shot.yml

Download data

To download xnli and pawsx run,

bash scripts/download_data.sh

Download Pretrained Models

Model | Description | Dataset | Checkpoints ---|---|---|--- XLMR-R large | Full model Finetuning with english data | XNLI | Please Contact 1/2 XLMR-R large | Full model Finetuning with english data | PAWSX | Please Contact 1/2 Inside the project create a folder named dumped. mkdir -p dumped Move the downloaded Pretrained Models to dumped.

mv pawsx-xlmr-baseline-fp16 dumped/ mv xnli-xlmr-baseline-fp16 dumped/

Run experiment

For re-producing Table 1 results,

bash scripts/exp_scripts/xnli/FewShotBenchmark/xlmr-baseline-few-shot-benchmark.sh

For re-producing Table 2 results,

bash scripts/exp_scripts/pawsx/FewShotBenchmark/xlmr-baseline-few-shot-benchmark.sh

for re-producing Table 3 results,

bash scripts/exp_scripts/pawsx/FewShotBenchmark/xlmr-baseline-cross-task-xnli-benchmark.sh

if you want to start evaluation of multiple seed models in multiple GPU, you can do that by following,

bash scripts/run.sh

Accumulate result

You can accumulate results by,

python scripts/extract_answer.py --folder_path dumped/xnli-xlmr-baseline-cross-lingual-transfer --shot 5 --lang en es --pkl_res_file_name "task-xnli-src_lang-en-tgt_lang-{}-lr_rate-0.0000075-shot-{}-seed-{}" python scripts/extract_answer.py --folder_path dumped/-xlmr-baseline-cross-lingual-transfer --shot 5 --lang "en" "de" "fr" --pkl_res_file_name "task-pawsx-src_lang-en-tgt_lang-{}-lr_rate-0.0000075-shot-{}-seed-{}" python scripts/extract_answer.py --folder_path dumped/pawsx-xlmr-cross-task-fewshot-benchmark --shot 5 --lang "en" "de" "fr" --pkl_res_file_name "task-pawsx-src_lang-en-tgt_lang-{}-lr_rate-0.0000075-shot-{}-seed-{}"

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.

Citation

When using this repository, please cite the following work:

bibtex @misc{bari2021nearest, title={Nearest Neighbour Few-Shot Learning for Cross-lingual Classification}, author={M Saiful Bari and Batool Haider and Saab Mansour}, year={2021}, eprint={2109.02221}, archivePrefix={arXiv}, primaryClass={cs.CL} }