Awesome

DEGREE: A Data-Efficient Generation-Based Event Extraction Model

Code for our NAACL-2022 paper DEGREE: A Data-Efficient Generation-Based Event Extraction Model.

Environment

• Python==3.8
• PyTorch==1.8.0
• transformers==3.1.0
• protobuf==3.17.3
• tensorboardx==2.4
• lxml==4.6.3
• beautifulsoup4==4.9.3
• bs4==0.0.1
• stanza==1.2
• sentencepiece==0.1.95
• ipdb==0.13.9

Note:

• If you meet issues reated to rust when installing transformers through pip, this website might be helpful

• Or you can reference the env_reference.yml for clearer installation

Datasets

We support ace05e, ace05ep, and ere.

Preprocessing

Our preprocessing mainly adapts OneIE's released scripts with minor modifications. We deeply thank the contribution from the authors of the paper.

ace05e

1. Prepare data processed from DyGIE++
2. Put the processed data into the folder processed_data/ace05e_dygieppformat
3. Run ./scripts/process_ace05e.sh

ace05ep

1. Download ACE data from LDC
2. Run ./scripts/process_ace05ep.sh

ere

1. Download ERE English data from LDC, specifically, "LDC2015E29_DEFT_Rich_ERE_English_Training_Annotation_V2", "LDC2015E68_DEFT_Rich_ERE_English_Training_Annotation_R2_V2", "LDC2015E78_DEFT_Rich_ERE_Chinese_and_English_Parallel_Annotation_V2"
2. Collect all these data under a directory with such setup:

ERE
├── LDC2015E29_DEFT_Rich_ERE_English_Training_Annotation_V2
│     ├── data
│     ├── docs
│     └── ...
├── LDC2015E68_DEFT_Rich_ERE_English_Training_Annotation_R2_V2
│     ├── data
│     ├── docs
│     └── ...
└── LDC2015E78_DEFT_Rich_ERE_Chinese_and_English_Parallel_Annotation_V2
      ├── data
      ├── docs
      └── ...

3. Run ./scripts/process_ere.sh

The above scripts will generate processed data (including the full training set and the low-resourece sets) in ./process_data.

Training

DEGREE (End2end)

Run ./scripts/train_degree_e2e.sh or use the following commands:

Generate data for DEGREE (End2end)

python degree/generate_data_degree_e2e.py -c config/config_degree_e2e_ace05e.json

Train DEGREE (End2end)

python degree/train_degree_e2e.py -c config/config_degree_e2e_ace05e.json

The model will be stored at ./output/degree_e2e_ace05e/[timestamp]/best_model.mdl in default.

DEGREE (ED)

Run ./scripts/train_degree_ed.sh or use the following commands:

Generate data for DEGREE (ED)

python degree/generate_data_degree_ed.py -c config/config_degree_ed_ace05e.json

Train DEGREE (ED)

python degree/train_degree_ed.py -c config/config_degree_ed_ace05e.json

The model will be stored at ./output/degree_ed_ace05e/[timestamp]/best_model.mdl in default.

DEGREE (EAE)

Run ./scripts/train_degree_eae.sh or use the following commands:

Generate data for DEGREE (EAE)

python degree/generate_data_degree_eae.py -c config/config_degree_eae_ace05e.json

Train DEGREE (EAE)

python degree/train_degree_eae.py -c config/config_degree_eae_ace05e.json

The model will be stored at ./output/degree_eae_ace05e/[timestamp]/best_model.mdl in default.

Evaluation

Evaluate DEGREE (End2end) on Event Extraction task

python degree/eval_end2endEE.py -c config/config_degree_e2e_ace05e.json -e [e2e_model]

Evaluate DEGREE (Pipe) on Event Extraction task

python degree/eval_pipelineEE.py -ced config/config_degree_ed_ace05e.json -ceae config/config_degree_eae_ace05e.json -ed [ed_model] -eae [eae_model]

Evaluate DEGREE (EAE) on Event Argument Extraction task (given gold triggers)

python degree/eval_pipelineEE.py -ceae config/config_degree_eae_ace05e.json -eae [eae_model] -g

Pre-Trained Models

Citation

If you find that the code is useful in your research, please consider citing our paper.

@inproceedings{naacl2022degree,
    author    = {I-Hung Hsu and Kuan-Hao Huang and Elizabeth Boschee and Scott Miller and Prem Natarajan and Kai-Wei Chang and Nanyun Peng},
    title     = {DEGREE: A Data-Efficient Generative Event Extraction Model},
    booktitle = {Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)},
    year      = {2022},
}

Contact

If you have any issue, please contact I-Hung Hsu at (ihunghsu@usc.edu) or Kuan-Hao Huang at (khhuang@cs.ucla.edu).