Nemo: Guiding and Contextualizing Weak Supervision for Interactive Data Programming

Code for paper Nemo: Guiding and Contextualizing Weak Supervision for Interactive Data Programming (VLDB 2023)

Environment Setup

python3 -m venv nemo_venv
source nemo_venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
pip uninstall -y nvidia_cublas_cu11

Download Data

gdown 1C48r6FCw-hU6ACbO9BMIgdtHkJn6AMB2
unzip nemo_data.zip && rm nemo_data.zip

Example Command Usages

Under directory src/:

• Snorkel (select by random):

  python interactive_dp.py  --dataset AmazonReview  --label-model snorkel --soft-training --query-method random
  

• Snorkel-Abs (select by abstain):

  python interactive_dp.py  --dataset AmazonReview  --label-model snorkel --soft-training --query-method abstain
  

• Snorkel-Dis (select by disagreement):

  python interactive_dp.py  --dataset AmazonReview  --label-model snorkel --soft-training --query-method disagreement
  

• Nemo:

  python interactive_dp.py  --dataset AmazonReview  --label-model snorkel --soft-training --query-method uncertainty_lm --seu --aggregate weighted --discard grid
  

Cite

If you find this repository useful, please consider citing:

@article{hsieh2022nemo,
  title={Nemo: Guiding and contextualizing weak supervision for interactive data programming},
  author={Hsieh, Cheng-Yu and Zhang, Jieyu and Ratner, Alexander},
  journal={arXiv preprint arXiv:2203.01382},
  year={2022}
}