UGDRE

Code for ACL 2023 paper Uncertainty Guided Label Denoising for Document-level Distant Relation Extraction.

Installation

  conda env create -f environment.yml
  conda activate UGDRE

Dataset

We perform experiments on DocRED and RE-DocRED.

Our Denoised data

For the DocRED dataset, our denoised data can be found at this link. For the RE-DocRED dataset, our denoised data can be found at this link.

Training and Evaluation

Pretrain the DRE model with DS data:

  bash scripts/run_pretrain.sh

Fine-tune the DRE model with human-annotated data:

  bash scripts/run_finetune.sh

Generate pseudo instances with uncertainty scores

  bash scripts/generate_pseudo_uncertainty.sh

Perform a re-label strategy to obtain denoised DS data

  python dataset.py

Part of the code is adapted from ATLOP.

Citation

@article{sun2023uncertainty,
  title={Uncertainty Guided Label Denoising for Document-level Distant Relation Extraction},
  author={Sun, Qi and Huang, Kun and Yang, Xiaocui and Hong, Pengfei and Zhang, Kun and Poria, Soujanya},
  journal={Proceedings of ACL},
  year={2023}
}