This repository provides the implementation for "Preserving Generalization of Language Models in Few-shot Continual Relation Extraction." Our approach leverages underutilized language model heads with mutual information maximization to maintain pre-trained knowledge and enhance few-shot continual learning for relation extraction.

To run the code, please install the following dependencies:

transformers==4.20.0
wordninja
wandb
scikit-learn
tqdm
numpy==1.23.0
peft
accelerate
sentencepiece
protobuf

We perform experiments using two publicly available relation extraction datasets:

• FewRel: A large-scale few-shot relation extraction dataset.
• TACRED: A widely-used dataset for relation classification.

To train BERT models on TACRED with 5-shot settings, follow these steps:

1. Navigate to the CPL scripts directory:

   cd CPL/bash

2. Run the 5-shot training script:

   bash tacred_5shot.sh

Alternatively, you can directly run the training script for different components:

1. For SCKD with Mutual Information Maximization (MMI):

   cd SCKD
   python main-mmi.py --task tacred --shot 5 

2. For ConPL:

   cd ConPL
   python main.py --task tacred --shot 5  

To train BERT models on FewRel with 5-shot settings:

1. Run the 5-shot script from the CPL directory:

   cd CPL/bash
   bash fewrel_5shot.sh

2. Alternatively, run the training commands directly:

   • SCKD with MMI:

     cd SCKD
     python main-mmi.py --task fewrel --shot 5 

   • ConPL:

     cd ConPL
     python main.py --task fewrel --shot 5  

To train using LLAMA2, ensure that you have set up your Hugging Face token (hf_token) in the required scripts:

• For ConPL, add the token in main-llm.py and dataprocess.py.
• For SCKD, add the token in sampler.py, main-llm.py, and main-llm-mmi.py.

1. To run SCKD with MMI:

   cd SCKD
   python main-llm-mmi.py --task tacred --shot 5 

2. To run ConPL:

   cd ConPL
   python main-llm.py --task tacred --shot 5  

1. To run SCKD with MMI:

   cd SCKD
   python main-llm-mmi.py --task fewrel --shot 5 

2. To run ConPL:

   cd ConPL
   python main-llm.py --task fewrel --shot 5