Awesome

OHRE

Dataset and code for NAACL 2021 paper: Open Hierarchical Relation Extraction.

Required packages:

torch==1.3.0.post2
torchsummary==1.5.1
metrics==0.3.3
numpy==1.16.2
torchvision==0.4.2
scikit-learn==0.20.3
python-louvain==0.13
matplotlib==3.0.3

OR install with:

pip install -r requirements.txt

Data

FewRel data and the preprocessing code are under the directory ./data.

NYT-FB data is from Discrete-State Variational Autoencoders for Joint Discovery and Factorization of Relations. (If you get the access to the original NYT-FB data from the authors, contact drogozhang@gmail.com to get our version of NYT-FB data.)

Run:

OpenRE Setting(e.g. on FewRel Hierarchy)

• with dynamic triplet loss (default):

python train_OHRE.py --dataset ori --gpu 0

• with virtual adversarial training:

python train_OHRE.py --dataset ori --gpu 0 --trainset_loss_type triplet_v_adv

Hierarchy Expansion Setting (e.g. on FewRel Hierarchy)

• with dynamic triplet loss (default):

python train_OHRE_hierarchy_eval_louvain.py --dataset ori --gpu 0

• with virtual adversarial training:

python train_OHRE_hierarchy_eval_louvain.py --dataset ori --gpu 0 --trainset_loss_type triplet_v_adv

Cite

If you use the dataset or the code, please cite this paper:

@inproceedings{zhang2021Open,
  title={Open Hierarchical Relation Extraction},
  author={Kai Zhang, Yuan Yao, Ruobing Xie, Xu Han, Zhiyuan Liu, Fen Lin, Leyu Lin, Maosong Sun},
  booktitle={Proceedings of NAACL 2021},
  year={2021}
}

Question

If you have any questions, please feel free to contact drogozhang@gmail.com.