/STA

Code for the paper "STA: Self-controlled Text Augmentation for Improving Text Classifications" (http://arxiv.org/abs/2302.12784)

Primary LanguagePython

The code for the paper "STA: Self-controlled Text Augmentation for Improving Text Classification"

To run this code, make sure you have

pip install transformers==4.10.0

The datasets in the work include SST-2, EMOTION, TREC and HumAID. They should be put in ./data/ (they are also downloadable from https://huggingface.co/datasets. For HumAID, the dataset can be downloaded from: https://crisisnlp.qcri.org/humaid_dataset).

To run STA on each dataset, simply execute the following commands

nohup python train_sst2.py > train_sst2_self_2021-2030.out &
nohup python train_emotion.py > train_emotion_self_2021-2030.out &
nohup python train_trec.py > train_trec_self_2021-2030.out &
nohup python train_humaid.py > train_humaid_self_2021-2030.out &

To run other variants such as STA-noself, just change the variable self_control from True to False in the corresponding scripts.

  • It is expected to obtain better results using larger generation and downstream models such as roberta-large and t5-large in the scripts

To cite this paper

@article{wang2023sta,
  title={STA: Self-controlled Text Augmentation for Improving Text Classification},
  author={Congcong Wang, Gonzalo Fiz Pontiveros, Steven Derby and Tri Kurniawan Wijaya},
  journal={arXiv preprint arXiv:2302.12784},
  year={2023}
}