DSGAN: Generative Adversarial Training for Distant Supervision Relation Extraction

PyTorch implementation of the Adversarial Training for Distant Supervision Relation Extraction model described in our ACL 2018 paper DSGAN: Generative Adversarial Training for Distant Supervision Relation Extraction. In this work, we define a Generator to find the true positive instances for relation extraction, and a Discriminator to do adversarial learning to robust the Generator. The core idea is assigning opposite labels to the instances predicted by Generator, to train the Discriminator; the optimal epoch is obtained until the performance of Discrinimator has the largest drop.

Steps to run the experiments

Requirements

• Python 2.7.12
• PyTorch 0.4.1
• panda 0.19.1
• sklearn 0.19.1

Datasets and word embeddings

• [Dataset and Pretrained word embeddings] are from OpenNRE. Please download(Baidu Yun or Google Drive) and put in into this directory.
• We include two versions of training dataset; they have different size, 522611 sentences and 570088 sentences repectively. This two options are included in args.py. Compared with 570088 version, 522611 version removes entity pairs that are repetitive with test dataset. 522611 is the default options in args.py.

Training

• python train.py

Output

• The cleaned dataset is outputed to the directory ./cleaned_data.

Test

• In order to validate the performance, we run thunlp/NRE on the cleaned dataset. For convenience, we have put their code in to the directory ./NRE-master.
• Taking CNN-ONE model as an example, run the code by
• make
• ./train
• The Precision-Recall file is outputed to ./NRE-master/CNN-ONE/out. Good Precision-Recal curves can be obtained from pr11.txt to pr14.txt.

Plot

• plot_PR_curve.ipynb

Reference

@article{qin2018dsgan,
  title={DSGAN: Generative Adversarial Training for Distant Supervision Relation Extraction},
  author={Qin, Pengda and Xu, Weiran and Wang, William Yang},
  journal={arXiv preprint arXiv:1805.09929},
  year={2018}
}