PRAGAS: Enhancing Few-Shot Continual Relation Extraction via Pairwise Augmentation and Guided Sharpness-Aware Optimization

Project Structure

PRAGAS/
├── FCRE/              # Few-Shot Continual Relation Extraction
│   ├── CPL/           # Contrastive Prompt Learning
│   └── SIRUS/         # Similar Relation Clusters
└── CED/               # Continual Event Detection
    └── SharpSeq/      # Sharp Sequence Learning

---

Few-Shot Continual Relation Extraction (FCRE)

Setup

1. Change directory

cd FCRE

2. Install requirements

pip install -r requirements.txt

Running Experiments

CPL (Contrastive Prompt Learning)

With BERT encoder:

cd CPL
bash bash/bash.sh

With Large Language Models (LLM):

cd CPL
bash bash/bash_llm.sh

SIRUS Baseline

With BERT encoder:

cd SIRUS/BERT
bash bash/fewrel_5shot.sh    # For FewRel dataset
bash bash/tacred_5shot.sh    # For TACRED dataset

With Large Language Models (LLM):

cd SIRUS/LLM
bash bash/fewrel_5shot.sh    # For FewRel dataset
bash bash/tacred_5shot.sh    # For TACRED dataset

---

Continual Event Detection (CED)

Setup

1. Change directory

cd CED

2. Install requirements

pip install -r requirements.txt

Running Experiments

SharpSeq on MAVEN dataset:

bash sh/maven.sh

---

Citation

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

@article{pragas2025,
    title={PRAGAS: Enhancing Few-Shot Continual Relation Extraction via Pairwise Augmentation and Guided Sharpness-Aware Optimization},
    author={Dat, Hoang Thanh and Anh, Nguyen Hoang and Hai, Nam Le and Van, Linh Ngo and Dinh, Sang},
    journal={Artificial Intelligence},
    year={2025},
    note={Manuscript Number: ARTINT-D-25-01408}
}