Continual Few-Shot Learning for Text Classification (EMNLP 2021)

Overview

This repository contains the datasets that we curated as a benchmark for Continual Few-Shot Learning (CFL) task. In CFL task, a system is challenged with a difficult phenomenon and asked to learn to correct mistakes with only a few (10 to 15) training examples.

Datasets Format

You can download the datasets from here.

In our CFL framework we consider two scenarios: (1) few-shot learning setup; (2) continual few-shot learning setup. In each setup, we have 5 and 2 categories for SNLI and IMDB datasets, respectively. For few-shot learning setup, each category has 3 train sets. Below is an overview of the directory structure.

.
├── few_shot                   
│   ├── SNLI        
│   │    ├── change-action-1
│   │    │    ├── (train/dev/test).tsv
│   │    ├── change-action-2
│   │    │    ├── (train/dev/test).tsv
│   │    ├── change-action-3
│   │    │    ├── (train/dev/test).tsv
│   ├── IMDB
│   │    ├── ...
├── continual
│   ├── SNLI        
│   │    ├── change-action
│   │    │    ├── (train/dev/test).tsv
└── ...

Categories

Dataset	Categories
SNLI	Insert/remove phrases, Substitute entities, Substitute evidence, Modify entity details, and Change action
IMDB	Modifiers and Negation

Reference

If you find this code helpful, please consider citing the following paper:

@inproceedings{pasunuru2021continual,
    title={Continual Few-Shot Learning for Text Classification},
    author={Pasunuru, Ramakanth and Stoyanov, Veselin and Bansal, Mohit},
    booktitle={EMNLP},
    year={2021}
}