EXCEEDS: Extracting Complex Events via Nugget-based Grid Modeling in Scientific Domain

This is the repository for the paper EXCEEDS: Extracting Complex Events via Nugget-based Grid Modeling in Scientific Domain.

🔥 News

• 2026 May 18: Paper is selected as an oral paper by ACL 2026 committee. Welcome to meet us in San Diego! The oral will be held at Harbor G, Session 2, Oral Session A: Information Extraction and Retrieval 1, on Sun. July 5, 11:00-12:30.
• 2026 April 28: Paper is updated on arXiv.
• 2026 April 24: Dataset is updated on HuggingFace.
• 2026 April 7: Paper is accepted by ACL 2026 Main Conference.
• 2025 Nov 11: Dataset is released on HuggingFace.
• 2024 Jun 20: Paper is available on arXiv.

📊 Release of SciEvents Dataset

You can find the released dataset in this HuggingFace repository.

The data format of SciEvents can be found at data/SciEvents/README.md.

🔁 Reproduction of EXCEEDS

Dataset and Pre-trained Model

Download SciEvents from this HuggingFace repository. Put train.json, dev.json, and test.json in data/SciEvents/ dicectory.

Download Roberta-large from this HuggingFace repository.

Environment

conda create -n exceeds python=3.8 -y
conda activate exceeds

# use CUDA 11.8 for example, check your own cuda version.
pip install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118

pip install \
  "transformers==4.30.0" \
  "numpy>=1.24,<2" \
  "tqdm>=4.65" \
  "prettytable>=3.7"

Train

python main.py --config config/scievents.json --output_dir outputs

We provide default arguments, which can be found in main.py and config/scievents.json.

Predict

python main.py --config config/scievents.json --ckpt your_best_model.state

We provide a checkpoint and its training log, which can be found in this HuggingFace repository.

📎 Citation

If you find this repository useful for your research, please cite our paper:

@misc{lu2026exceedsextractingcomplexevents,
      title={EXCEEDS: Extracting Complex Events via Nugget-based Grid Modeling in Scientific Domain}, 
      author={Yi-Fan Lu and Xian-Ling Mao and Bo Wang and Xiao Liu and Heyan Huang},
      year={2026},
      eprint={2406.14075},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.14075}, 
}