Prototype Conditioned Generative Replay for Continual Learning in NLP

This repository contains code for the paper "Prototype Conditioned Generative Replay for Continual Learning in NLP". We propose Prototype Conditioned Generative Replay (PCGR), which enhances generative reply by incorporating prototype, the task-level statistics. PCGR leverages a Prototype Conditioned Variational Autoencoder (PCVAE) to generate pseudo-samples that are both representative and diverse, accurately reflecting the real data distribution. This approach can effectively mitigates catastrophic forgetting in continual learning.

Data Preparation

1. Define the sequential order of your tasks from which your model will continuously learn. Place the corresponding information in the file ./SnD/info.py using the following format:

   ORDER_DICT = {   
       'order_0': ['example_task_0', 'example_task_1'],
   }
   

2. For each task, tasking example_task_1 as an example (which is a subset of the intent detection task: "banking77"), follow these steps:

   i). Organize your data in the ./datasets folder with the following structure:

   ./datasets/
   --example_task_0/
       --train.json
       --valid.json
       --test.json
   

   ii). Write the corresponding "Dataset" class, place it under ./SnD/SnD_datasets, and import the dataset in the file ./SnD/SnD_models/trainer.py. Refer to ./SnD/SnD_datasets/cls_dataset.py as an example.

   iii). Add the corresponding task-specific metric in the file ./SnD/metrics.py, and include an entry for it in the compute_metrics function. Refer to compute_em function as an example.

   iv). Add the relevant task-specific information into the file ./SnD/info.py, following this format:

   TASK2INFO = {   
           'example_task_0': {   
                   'dataset_class': 'PromptCLSDataset', 
                   'dataset_folder': 'example_task_0',  
                   'eval_metric': {'em'},               
                   'id': 0,                             
                   'metric': 'em',                      
                   'part_prompt': 'which intent category best describes:',
                   'task_type': 'intent'                
               },
           
       }
   IDTOTASK = {0: 'example_task_0',}
   
   TASKTOID = {'example_task_0': 0,}
   

If you want to use the exact dataset for training and testing PCGR, please email Xi Chen.

Environment and Setup

Experiments in this paper are run on Tesla V100 with CUDA 11.8

1. Install the experiment:

   pip install -r requirements.txt
   

2. Edit ./enviornment.

Training and Testing

sh ./scripts/pcgr.sh

Citation

@inproceedings{pcgr,
    title = "Prototype Conditioned Generative Replay for Continual Learning in {NLP}",
    author = "Chen, Xi and Zeng, Min",
    editor = "Chiruzzo, Luis and Ritter, Alan and Wang, Lu",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.636/",
    pages = "12754--12770",
    ISBN = "979-8-89176-189-6"
}

Thanks

The code of this repository partly relies on PCLL and I would like to show my sincere gratitude to authors of it.

Concact

If you have any questions about our paper and code, please contact Xi Chen via chenxi.mail.1005@gmail.com.

Note

You might find it confusing that our main code is located in the folder ./SnD, rather than something like PCGR. This is because our method focuses on improving pseudo sample quality in terms of semantic and distribution scales. In our initial draft, we named our method SnD to reflect this focus.