PEMSCL

This is the official implementation for "Towards Integration of Discriminability and Robustness for Document-Level Relation Extraction" (EACL 2023, Main Conference).

Citation

If you find our work helpful for your research, please cite our paper:

@inproceedings{pemscl,
    title = "Towards Integration of Discriminability and Robustness for Document-Level Relation Extraction",
    author = "Jia Guo and Stanley Kok and Lidong Bing",
    booktitle = "Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics",
    year = "2023"
}

Installation

>> conda create -n pemscl python==3.7.4
>> conda activate pemscl
>> conda install pytorch==1.7.0 torchvision cudatoolkit=10.2 -c pytorch

Other dependencies:

• tqdm==4.61.1
• transformers==4.8.2
• ujson==4.0.2
• numpy==1.19.4
• apex==0.1
• opt-einsum==3.3.0

The expected folder structure is:

PEMSCL
 |-- dataset
 |    |-- docred
 |    |    |-- train_annotated.json        
 |    |    |-- dev.json
 |    |    |-- test.json
 |    |-- re-docred
 |    |    |-- train_annotated.json        
 |    |    |-- dev.json
 |    |    |-- test.json
 |-- meta
 |    |-- rel2id.json
 |-- codes
 |-- scripts 
 |-- results
 |-- logs
 |-- checkpoints

Training models on benchmarks

>> cd scripts
>> sh run_docred.sh  # for the DocRED dataset
>> sh run_re-docred.sh # for the Re-DocRED dataset

Evaluation

You can evaluate your saved checkpoint by replacing the --save_path argument with the --load_path argument. Our trained models pemscl-docred and pemscl-re-docred can be found here.

Acknowledgement

We refer to the codes of ATLOP. Thanks for their contributions.