Text-AutoAugment (TAA)
This repository contains the code for our paper Text AutoAugment: Learning Compositional Augmentation Policy for Text Classification (EMNLP 2021 main conference).
Overview
-
We present a learnable and compositional framework for data augmentation. Our proposed algorithm automatically searches for the optimal compositional policy, which improves the diversity and quality of augmented samples.
-
In low-resource and class-imbalanced regimes of six benchmark datasets, TAA significantly improves the generalization ability of deep neural networks like BERT and effectively boosts text classification performance.
Getting Started
-
Prepare environment
conda create -n taa python=3.6 conda activate taa conda install pytorch torchvision cudatoolkit=10.0 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch pip install -r requirements.txt python -c "import nltk; nltk.download('wordnet'); nltk.download('averaged_perceptron_tagger')" -
Modify
datarootparameter inconfs/*yamlandabspathparameter inscript/*.sh:- e.g., change
dataroot: /home/renshuhuai/TextAutoAugment/data/aclImdbin confs/bert_imdb.yaml todataroot: path-to-your-TextAutoAugment/data/aclImdb - change
--abspath '/home/renshuhuai/TextAutoAugment'in script/imdb_lowresource.sh to--abspath 'path-to-your-TextAutoAugment'
- e.g., change
-
Search for the best augmentation policy, e.g., low-resource regime for IMDB:
sh script/imdb_lowresource.sh
scripts for policy search in the low-resource and class-imbalanced regime for all datasets are provided in the script/ fold.
-
Train a model with pre-searched policy in archive.py, e.g., train model in low-resource regime for IMDB:
python train.py -c confs/bert_imdb.yaml
train model on full dataset of IMDB:
python train.py -c confs/bert_imdb.yaml --train-npc -1 --valid-npc -1 --test-npc -1
Contact
If you have any questions related to the code or the paper, feel free to email Shuhuai (renshuhuai007 [AT] gmail [DOT] com).
Acknowledgments
Code refers to: fast-autoaugment.
Citation
If you find this code useful for your research, please consider citing:
@inproceedings{ren2021taa,
title={Text AutoAugment: Learning Compositional Augmentation Policy for Text Classification},
author={Shuhuai Ren, Jinchao Zhang, Lei Li, Xu Sun, Jie Zhou},
booktitle={EMNLP},
year={2021}
}
License
MIT
39 Jan 1, 2023
257 Dec 30, 2022
45 Jan 9, 2023
63 Nov 14, 2022
17 Jun 19, 2022
11 Jan 1, 2023
42 Jan 7, 2023
89 Dec 26, 2022
27 Oct 28, 2022
25 Sep 6, 2022
12 Nov 24, 2022
39 Dec 11, 2022
7 Oct 20, 2022
19 Aug 23, 2022
61 Jan 3, 2023
57 Jan 1, 2023
9 Oct 28, 2022