Distinct Label Representations for Few-shot Text Classification

Dataset

Dataset can be available on the repo of Bao et al.

Usage

# For run specific setting
python3 train.py --data huffpost --data_dir data/ --config_files config.yaml --lr 1e-5 --alpha 1e-4 

# For searching hyperparameters
python3 train.py --data huffpost --data_dir data/ --config_files config.yaml --search_params

# For evaluating
python3 train.py --data huffpost --data_dir data/ --config_file config.yaml --eval_model_path /path/to/model --test_N 5 --test_K 1

Config.yaml

name	description
model	The name of a model. Please set the value one of following: ["proto", "mlman", "maml"]
num_train_steps	The maximum number of training steps
num_eval_steps	The number of evaluation steps
early_stop	The threshold for early stopping. If the model cannot beat the best socre for early_stop epochs, training is stopped.
eval_interval	a model will be evaluated every eval_interval steps
meta_setting["K"]	The number of examples for each label
meta_setting["N"]	The number of labels
meta_setting["Q"]	The number of elements in query set

Citation

@inproceedings{ohashi-etal-2021-distinct,
    title = "Distinct Label Representations for Few-Shot Text Classification",
    author = "Ohashi, Sora  and
      Takayama, Junya  and
      Kajiwara, Tomoyuki  and
      Arase, Yuki",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-short.105",
    doi = "10.18653/v1/2021.acl-short.105",
    pages = "831--836",
    abstract = "Few-shot text classification aims to classify inputs whose label has only a few examples. Previous studies overlooked the semantic relevance between label representations. Therefore, they are easily confused by labels that are relevant. To address this problem, we propose a method that generates distinct label representations that embed information specific to each label. Our method is applicable to conventional few-shot classification models. Experimental results show that our method significantly improved the performance of few-shot text classification across models and datasets.",
}