NGS

Source code for AACL-IJCNLP 2020 paper "Neural Gibbs Sampling for Joint Event Argument Extraction".

Requirements

• python == 3.6.9
• pytorch == 0.6.1
• numpy == 1.15.2
• sklearn == 0.20.2
• pytorch-pretrained-bert == 0.4.0
• nltk
• tqdm

We use the ACE2005 (LDC2006T06) and TAC KBP 2016 (LDC2017E05) as our benchmarks. Due to the LDC license limitation, we cannot share the datasets.

For NGS (CNN), the 100-dim Glove word vectors pre-trained with Wikipedia 2014+Gigaword 5 is used.

Usage

NGS (CNN)

The codes are in the NGS-DMCNN folder.

1. run input.py to preprocess the data.
2. run trigger.py and argument.py to train and test the prior models for the ED and EAE.
3. run conditional.py to train and test the conditional neural model.
4. run Gibbs_an.py to run the Gibbs sampling + Simulated annealing.
5. hyper-parameters and data paths are specified in constant.py.

NGS (BERT)

The codes are in the NGS-DMBERT folder.

1. run input_bert_role.py to preprocess the data.
2. run trigger_bert.py and argument_bert.py to train and test the prior models for the ED and EAE.
3. run conditional_bert.py to train and test the conditional neural model.
4. run Gibbs_an_bert.py to run the Gibbs sampling + Simulated annealing.
5. hyper-parameters and data paths are specified in constant.py.

Cite

If the codes help you, please cite the following paper:

Neural Gibbs Sampling for Joint Event Argument Extraction. Xiaozhi Wang, Shengyu Jia, Xu Han, Zhiyuan Liu, Juanzi Li, Peng Li, Jie Zhou. AACL-IJCNLP 2020.