LDNet

Label Drop for Multi-Aspect Relation Modeling in Universal Information Extraction

Requirements

• Environment
  • Python >= 3.10
  • CUDA 11.8
  • Download deberta-v3-base model from Huggingface and put it under file microsoft/deberta-v3-large.
  • Run the command below:

  pip install -r requirements.txt
  

  • Install REx, and change REx/rex/tasks with tasks in this repo:

  git clone https://github.com/Spico197/REx.git
  cd REx
  pip install -e .
  

Quick Start

Pretrained Model weights & Datasets

Download pretrained model from Downton/LDNet_Pretrain and put the folder under LDNet_outputs folder. Download datasets from Mirror and put them under resources folder. Zero-shot datasets are in Spico/Mirror, put them under resources/Mirror folder. Download MIE datasets (Twitter-2015, Twitter-2017, MNRE) and transform them:

python data/txt2json.py ./ data/newInst/ T F F

Evaluation

# main tasks
bash scripts/single_task_wPTAllExcluded_wInstruction/run.sh

# Multi-span and N-ary extraction
bash scripts/single_task_wPTAllExcluded_wInstruction/run_new_tasks.sh

# Few-shot
bash scripts/single_task_wPTAllExcluded_wInstruction/fewshot.sh

# Zero-shot
python -m src.eval

# Ablation
bash scripts/ablation.sh

# MIE
bash MIE.sh

Training

rex train -m src.task -dc conf/Pretrain_ld.yaml

Citation

If you find our model/code/paper helpful, please consider cite our paper 📝 and star us ⭐️！

@misc{yang2025labeldropmultiaspectrelation,
      title={Label Drop for Multi-Aspect Relation Modeling in Universal Information Extraction}, 
      author={Lu Yang and Jiajia Li and En Ci and Lefei Zhang and Zuchao Li and Ping Wang},
      year={2025},
      eprint={2502.12614},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.12614}, 
}