xinyadu / doc_event_role

Document-level Event Role Filler Extraction (ACL 2020)

paper link

Please also check our sibling project on event entity extraction for template filling

Dependencies

• python 3.5.6
• spacy==2.0.12
• torch 0.4.1 (for ./model)
• pytorch-pretrained-bert==0.6.2 (for ./model)

Dataset

./data/
├── process_train_dev.py  # proc script
├── process_test.py       # proc script
│ 
├── processed/            # processed data files
│   ├── train.json 
│   ├── dev.json       
│   └── test.json           
│ 
└── raw_muc/              # Raw data files from MUC-{3,4}

Run preprocessing for train and dev, use flag -full to include all the templates.

python process_train_dev.py

Run preprocessing for test,

python process_test.py

Evaluation

To run the eval script:

python eval.py --goldfile <gold file path> --predfile <pred file path>

We use ./data/processed/test.json for <gold file path> in the experiments. We also include an example output file (./model/pred.json) in the model foler:

python eval.py --goldfile ./data/processed/test.json --predfile ./model/pred.json

If you use our eval script, please make sure the <pred file> is of the same format as pred.json.

Model Code

We also include a sample output file in the folder.

Citation

If you use materials in this repo helpful, please cite:

@inproceedings{du2020doucment,
    title={Document-Level Event Role Filler Extraction Using Multi-Granularity Contextualized Encoding of the Text},
    author={Du, Xinya and Cardie, Claire},
    booktitle={Association for Computational Linguistics (ACL)},
    year={2020}
  }