CurriculumDocRE: A Curriculum Learning Approach for DocRE

This repository contains the implementation of CurriculumDocRE, a curriculum learning framework for document-level relation extraction (DocRE). The model progressively learns from easy to hard relations based on a multi‑faceted complexity score that considers sentence distance, coreference resolution, and relation rarity.

Requirements

transformers>=4.14.1
numpy>=1.20.0
scikit-learn>=1.0.0
tqdm>=4.62.0
ujson
wandb
opt-einsum

Install with: pip install -r requirements.txt

Directory Structure

CurriculumDocRE/
├── README.md
├── requirements.txt
├── run.py
├── train.py
├── config.py
├── utils.py
├── prepro.py
├── model.py
├── losses.py
├── long_seq.py
├── data_augmentation.py
├── evaluation.py
├── args.py
├── scripts/
│   ├── run_roberta_cur.sh
│   └── run_bert_cur.sh
└── data/ (you need to download Re‑DocRED, CDR, GDA yourself)

Data Preparation

1. Download the required datasets:
   • Re‑DocRED: https://github.com/tonytan48/Re-DocRED
   • RE2‑DocRED: https://github.com/klassessg/re2-docred

The CDR and GDA datasets can be obtained following the instructions in edge-oriented graph.

2. Place the files in the data/ directory with the following structure:

data/
├── redocred/
│ ├── train.json
│ ├── dev.json
│ └── test.json
├── re2docred/ (optional)
├── cdr/
│ ├── train.json
│ ├── dev.json
│ └── test.json
└── gda/
├── train.json
├── dev.json
└── test.json

3. Also place rel2id.json and rel_info.json (provided in the DocRED repository) in a meta/ folder.

Running the Code

Training

Fully‑supervised baseline (no curriculum)

python -m src.run --do_train --data_dir data/redocred --train_file train.json --dev_file dev.json --test_file test.json --num_train_epochs 30

Curriculum learning (three stages)

python -m src.run --do_train --curriculum --phase1_epochs 10 --phase2_epochs 10 --phase3_epochs 10 --max_alpha 2.0 --data_dir data/redocred

With augmentation for rare relations

python -m src.run --do_train --curriculum --augment --augment_factor 1 --data_dir data/redocred

Training on biomedical datasets (binary)

CDR

python -m src.run --do_train --curriculum --num_labels 2 --num_class 2 --evi_lambda 0 --data_dir data/cdr --transformer_type bert --model_name_or_path bert-base-uncased

GDA

python -m src.run --do_train --curriculum --num_labels 2 --num_class 2 --evi_lambda 0 --data_dir data/gda --train_batch_size 2 --gradient_accumulation_steps 4 --transformer_type bert --model_name_or_path bert-base-uncased

Evaluation

On development set

python -m src.run --load_path ./output/curriculum_redocred/best.ckpt --data_dir data/redocred --dev_file dev.json --evaluate

On test set

python -m src.run --load_path ./output/curriculum_redocred/best.ckpt --data_dir data/redocred --test_file test.json --evaluate --test_only

Configuration

Argument	Default	Description
--curriculum	False	Enable curriculum learning
--phase1_epochs	10	Epochs for Stage 1 (easy relations, distance ≤1)
--phase2_epochs	10	Epochs for Stage 2 (medium relations, distance 2‑4)
--phase3_epochs	10	Epochs for Stage 3 (hard relations, distance >4)
--max_alpha	2.0	Exponent for curriculum weighting (higher = more focus on hard pairs)
--augment	False	Enable data augmentation for rare relations
--augment_factor	1	Number of augmented copies per positive document
--pos_weight	20	Positive class weight for ATLoss
--evi_lambda	0.5	Weight for evidence loss (if evidence annotations are available)
--learning_rate	3e-5 (RoBERTa) / 5e-5 (BERT)	Learning rate
--train_batch_size	4	Batch size per GPU
--gradient_accumulation_steps	2	Gradient accumulation steps
--max_seq_length	1024	Maximum input tokens
--warmup_ratio	0.06	Linear warmup proportion

Citation

If you use this code in your research, please cite:

@inproceedings{ali2026curriculum,
  title={CurriculumDocRE: A Curriculum Learning Approach for Document-Level Relation Extraction},
  author={Ali, Manzoor and Saleem, Muhammad and Ngonga Ngomo, Axel-Cyrille},
  year={2026}
}