This is the origin Pytorch implementation of EAREE in the following paper: Reinforcement Learning-based Dialogue Guided Event Extraction to Exploit Argument Relations, arXiv 2021
Figure 1. The architecture of EAREE.
- Python 3.6
- matplotlib == 3.1.1
- numpy == 1.19.4
- pandas == 0.25.1
- scikit-learn==0.21.3
- transformers==2.5.1
- torch == 1.4.0
Dependencies can be installed using the following command:
pip install -r requirements.txtPrepare ACE 2005 dataset.
(Download: https://catalog.ldc.upenn.edu/LDC2006T06. Note that ACE 2005 dataset is not free.)
The required data files should be put into dataset/ folder.
Download stanford-corenlp model.
wget http://nlp.stanford.edu/software/stanford-corenlp-full-2018-10-05.zip
unzip stanford-corenlp-full-2018-10-05.zipEvent extraction requires extracting all arguments and their roles corresponding to each event. Doing so is challenging because an event is often associated with more than one argument whose role can vary in different contexts.
Figure 2. An example on ACE 2005.
# Settings
1) Event Classification: TC\pybert\configs\basic_config.py
2) Argument Extraction: AI_RC\config\args.py
# Event Classification
1) Preprocess: python TC_preprocess.py
2) Data partition: python run_bert.py --do_data to preprocess data
3) Train: python run_bert.py --do_train --save_best
4) Test: python run_bert.py --do_test
# Argument Extraction
1) Preprocess: python AI_RC_preprocess.py
2) Data partition: python get_data.py
3) Train: python train_start.py
4) Test: python test.py
We evaluate event classification (EC), argument identification (AI), and role classification (RC) performance on the ACE dataset. The overall performance of our approach compared against the baseline methods is reported in Table 1, where the best results are shown in bold and “–” means not available.
Figure 3. Event classification, argument identification and argument role classification result on the ACE2005 test set.
To demonstrate the key designs of our model, we conduct ablation tests by evaluating four variants in Table 2. The result shows that the four variants are significantly valuable in EC, AI, and RC tasks.
Figure 4. Ablation Study on Global Constraints.