📝 GPT Self-Supervision for an Efficient Data Annotator

Core Idea: Iteratively swap the annotated objectives of (x, y); replace the template with one generated from data, which then self-refines based on the similarity score between cells.

Usage:

git clone https://github.com/TerryPei/Annotator.git
cd Annotator

Replace the API key in the prepare folder:

cd prepare
python encrypt_api_key.py

The APIs of OpenAI have been changed. When conducting the experiments, please reinstall the previous version '0.28.0':

pip install -U openai==0.28.0

One-Shot Stages:

python annotator.py

The correctly runing will output the log files like this:

Second stages:

Copy and paste the best template from the log file to generation python file, and run with

cd generation
python run_generate_summary.py

The recovery evaluation score could be directly get via eval.py.

The annotated dataset link https://drive.google.com/drive/folders/1MyxU1DE5hE5ANbDc3g8e4bSBVFQek5nH?usp=sharing.

Citation

If this paper helps your research, we would appreciate your citation:

@article{pei2023gpt,
  title={Gpt self-supervision for a better data annotator},
  author={Pei, Xiaohuan and Li, Yanxi and Xu, Chang},
  journal={arXiv preprint arXiv:2306.04349},
  year={2023}
}