This repository is the official implementation of DocuNet, which is model proposed in a paper: Document-level Relation Extraction as Semantic Segmentation, accepted by IJCAI2021 main conference.

• ❗NOTE: Docunet is integrated in the knowledge extraction toolkit DeepKE.

This paper innovatively proposes the DocuNet model, which first regards the document-level relation extraction as the semantic segmentation task in computer vision.

To install requirements:

pip install -r requirements.txt

To train the DocuNet model in the paper on the dataset DocRED, run this command:

>> bash scripts/run_docred.sh # use BERT/RoBERTa by setting --transformer-type

To train the DocuNet model in the paper on the dataset CDR and GDA, run this command:

>> bash scripts/run_cdr.sh  # for CDR
>> bash scripts/run_gda.sh  # for GDA

To evaluate the trained model in the paper, you setting the --load_path argument in training scripts. The program will log the result of evaluation automatically. And for DocRED it will generate a test file result.json in the official evaluation format. You can compress and submit it to Colab for the official test score.

Our model achieves the following performance on :

Part of our code is borrowed from https://github.com/wzhouad/ATLOP, many thanks. You can refer to https://github.com/fenchri/edge-oriented-graph for the detailed preprocessing process of GDA and CDR datasets (acquire the file of train_filter.data, dev_filter.data and test_filter.data).

If you use or extend our work, please cite the paper as follows:

@inproceedings{ijcai2021-551,
  title     = {Document-level Relation Extraction as Semantic Segmentation},
  author    = {Zhang, Ningyu and Chen, Xiang and Xie, Xin and Deng, Shumin and Tan, Chuanqi and Chen, Mosha and Huang, Fei and Si, Luo and Chen, Huajun},
  booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI-21}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Zhi-Hua Zhou},
  pages     = {3999--4006},
  year      = {2021},
  month     = {8},
  note      = {Main Track}
  doi       = {10.24963/ijcai.2021/551},
  url       = {https://doi.org/10.24963/ijcai.2021/551},
}