Prompt-based Data Augmentation for Low-Resource NLU Tasks

This repository is the official implementation of PromDA: Prompt-based Data Augmentation for Low-Resource NLU Tasks.

Requirements

To install requirements:

conda create --name exp --file requirements.txt

Pre-training Soft Prompt

To obtain C4 realnewslike split, please run:

python get_large_pre_training_c4_data.py

We need to do the Prompt Pre-training, please run:

python pre_train_t5.py --config model_config/pre_train_keyword_pt.yml --serialization-dir pretrain_web_page_keyword_t5_short --train

In a Nvidia A100 GPU, it takes about 24 hours to complete the pre-training.

We also provide the checkpoint that we used in the paper here. Please put the file folder pretrain_web_page_keyword_t5_short in the root directory of this project.

Run PromDA

To run the full data augmentation experiments, please follow below instructuins:

Sequence Labelling Tasks

To run the wikiann experiments on the first GPU under the shot-10 setting using random seed 18, please run

bash script/run_few_shot_bert_prefix.sh 0 18 10 wikiann 1000

Sentence Classification Tasks

To run the sst2 experiments on the first GPU under the shot-10 setting using random seed 18, please run