---

GEGA

This repository contains codes for the paper “GEGA: Graph Convolutional Networks and Evidence Retrieval Guided Attention for Enhanced Document-level Relation Extraction”.

🔥🔥🔥GEGA is accepted by ICASSP 2025(main conference)! Thanks to all collaborators！

Requirements

The following packages are required:

• Python (tested on 3.8.13)
• CUDA (tested on 11.6)
• PyTorch (tested on 1.11.0)
• Transformers (tested on 4.14.1)
• numpy (tested on 1.22.4)
• opt-einsum (tested on 3.3.0)
• wandb
• ujson
• tqdm

Training

1. Fully-supervised setting
   Run the following command for BERT training:

   bash scripts/run_bert.sh ${name} ${lambda} 0.05

2. Inference on distantly-supervised data
   Run the following command for BERT inference:

   bash scripts/infer_distant_bert.sh ${name} ${load_dir}

3. Self-Training on distantly-supervised data
   Run the following command for BERT self-training:

   bash scripts/run_self_train_bert.sh ${name} ${teacher_signal_dir} ${lambda} ${seed}

4. Fine-tuning on human-annotated data
   Run the following command for BERT fine-tuning:

   bash scripts/run_finetune_bert.sh ${name} ${student_model_dir} ${lambda} ${seed}

Evaluation

Run the following command for evaluation using BERT:

bash scripts/isf_bert.sh ${name} ${model_dir} test

Citation

If you make use of this code in your work, please kindly cite our paper:

@article{mao2024gega,
  title={GEGA: Graph Convolutional Networks and Evidence Retrieval Guided Attention for Enhanced Document-level Relation Extraction},
  author={Mao, Yanxu and Chen, Xiaohui and Liu, Peipei and Cui, Tiehan and Yue, Zuhui and Li, Zheng},
  journal={arXiv preprint arXiv:2407.21384},
  year={2024}
}