This code is originally released from amazon-research package (https://github.com/amazon-research/transformers-data-augmentation) In the paper, we mentioned https://github.com/varinf/TransformersDataAugmentation url so we are providing a copy of the same code here.

Code associated with the Data Augmentation using Pre-trained Transformer Models paper

Code contains implementation of the following data augmentation methods

• EDA (Baseline)
• Backtranslation (Baseline)
• CBERT (Baseline)
• BERT Prepend (Our paper)
• GPT-2 Prepend (Our paper)
• BART Prepend (Our paper)

In paper, we use three datasets from following resources

• STSA-2 : https://github.com/1024er/cbert_aug/tree/crayon/datasets/stsa.binary
• TREC : https://github.com/1024er/cbert_aug/tree/crayon/datasets/TREC
• SNIPS : https://github.com/MiuLab/SlotGated-SLU/tree/master/data/snips

Run src/utils/download_and_prepare_datasets.sh file to prepare all datsets.
download_and_prepare_datasets.sh performs following steps

1. Download data from github
2. Replace numeric labels with text for STSA-2 and TREC dataset
3. For a given dataset, creates 15 random splits of train and dev data.

To run this code, you need following dependencies

• Pytorch 1.5
• fairseq 0.9
• transformers 2.9

To run data augmentation experiment for a given dataset, run bash script in scripts folder. For example, to run data augmentation on snips dataset,

• run scripts/bart_snips_lower.sh for BART experiment
• run scripts/bert_snips_lower.sh for rest of the data augmentation methods

@inproceedings{kumar-etal-2020-data,
    title = "Data Augmentation using Pre-trained Transformer Models",
    author = "Kumar, Varun  and
      Choudhary, Ashutosh  and
      Cho, Eunah",
    booktitle = "Proceedings of the 2nd Workshop on Life-long Learning for Spoken Language Systems",
    month = dec,
    year = "2020",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.lifelongnlp-1.3",
    pages = "18--26",
}

Please reachout to [email protected] for any questions related to this code.

This project is licensed under the Creative Common Attribution Non-Commercial 4.0 license.