This repository contains the source code for SEDA. You can follow the instructions below to reproduce our experiments. Our source code references https://github.com/shichao-wang/CircEvent.git

Our experiments are conducted on the New York Times (NYT) portion of the English Gigawords. You can get access from the official website. The data split we used is provided by Granroth-Wilding[1] We annotate the raw documents based on Lee[2] with the standford CoreNLP toolkit. The configuration of CoreNLP is listed in corenlp.props file.

We conducted our experiments with on a workstation with a tesla V100. Our programs are tested under PyTorch 1.8.1 + CUDA 10.2.

1. Setup Python environment. We encourage using conda to setup the python virtual environment. conda create -n seda python==3.8 && conda activate seda
2. Install the CUDA toolkit and Pytorch. `pip install torch==1.13.0+cu116 torchvision==0.14.0+cu116 --extra-index-url https://download.pytorch.org/whl/cu116 -i https://pypi.tuna.tsinghua.edu.cn/simple``
3. Install the pip packages. pip install -r requirements.txt
4. Install the circumst_event package pip install -e .

Now the environment has been set up in the seda virtual environment.

The source codes can be divided into two parts, i.e. data preprocessing and model training. The entry scripts are placed in bin folder. Each step and its corresponding script is listed below.

1. extract text out of gigaword xml file. 1_extract_gigaword_nyt.py
2. annotate text with CoreNLP. 2_corenlp_annotate.py
3. extract event chain from annotated document. 3_extract_event_chain.py
4. convert event chain words to ids. 4_index_event_chain.py
5. split into train, validation, test set. 5_split_dataset.py
6. train the circ model. 6_circ_train.py
7. evaluate the saved model and generate quantitative analysis file. 7_circ_eval.py
8. draw figures of changing in accuracies based on quantitative analysis file. 8_mask_multiple_items_with_weights.py

[1] Shichao Wang, Xiangrui Cai, Hongbin Wang, and Xiaojie Yuan. 2021. Incorporating circumstances into narrative event prediction. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4840–4849.

[2] Mark Granroth-Wilding and Stephen Clark. 2016. What happens next? Event Prediction Using a Compositional Neural Network Model. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pages 2727–2733, Phoenix, Arizona, February. AAAI Press.

[3] I-Ta Lee and Dan Goldwasser. 2019. Multi-Relational Script Learning for Discourse Relations. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4214–4226, Florence, Italy, July. Association for Computational Linguistics.