GDAP

Code for Generating Disentangled Arguments with Prompts: A Simple Event Extraction Framework that Works

Environment

• Python (verified: v3.8)
• CUDA (verified: v11.1)
• Packages (see requirements.txt)

Usage

Preprocessing

We follow dygiepp for data preprocessing.

• text2et: Event Type Detection
• ettext2tri: Trigger Extraction
• etrttext2role: Argument Extraction

# data processed by dyieapp
data/text2target/dyiepp_ace1005_ettext2tri_subtype
├── event.schema 
├── test.json
├── train.json
└── val.json

# data processed by  data_convert.convert_text_to_target
data/text2target/dyiepp_ace1005_ettext2tri_subtype
├── event.schema
├── test.json
├── train.json
└── val.json

Useful commands:

python -m data_convert.convert_text_to_target # data/raw_data -> data/text2target
python convert_dyiepp_to_sentence.py data/raw_data/dyiepp_ace2005 # doc -> sentence, used in evaluation

Training

Relevant scripts:

• run_seq2seq.py: Python code entry, modified from the transformers/examples/seq2seq/run_seq2seq.py
• run_seq2seq_span.bash: Model training script logging to the log file.

Example (see the above two files for more details):

# ace05 event type detection t5-base, the metric_format use eval_trigger-F1 
bash run_seq2seq_span.bash --data=dyiepp_ace2005_text2et_subtype --model=t5-base --format=et --metric_format=eval_trigger-F1

# ace05 tri extraction t5-base
bash run_seq2seq_span.bash --data=dyiepp_ace2005_ettext2tri_subtype --model=t5-base --format=tri --metric_format=eval_trigger-F1

# ace05 argument extraction t5-base
bash run_seq2seq_span.bash --data=dyiepp_ace2005_etrttext2role_subtype --model=t5-base --format=role --metric_format=eval_role-F1

Trained models are saved in the models/ folder.

The event type detection use the same output format and metric_format as trigger extraction, so the et exp result is included in eval_trigger- and test_trigger- of the log.

Evaluation

• run_tri_predict.bash: trigger extraction evaluation and inference script.
• run_arg_predict.bash: argument extraction evaluation and inference script.

If you find this repo helpful...

Please give us a :star: and cite our paper as

@inproceedings{si2021-GDAP,
      title={Generating Disentangled Arguments with Prompts: A Simple Event Extraction Framework that Works}, 
      author={Jinghui Si and Xutan Peng and Chen Li and Haotian Xu and Jianxin Li},
      booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
      year={2022}
}

This project borrows code from Text2Event