FewOne

This directory is supplemented with the paper entitled "Classical Sequence Match is Competitive Few-Shot One-Class Learner"

Environment

Python: 3.7.13
Pytorch: 1.11.0
transformers: 4.14.1
sklearn: 1.0.2
numpy: 1.21.2

Datasets

The datasets directory contains the two datasets, ACD and Huffpost, in the experiments. ACD is too large. Here we only upload the Huffpost. Read dataset with the following code:

text_word_id, text_id_to_word = pickle.load(open(path + "/text_id_info.p", "rb"))
word_embedding = pickle.load(open(path + "/embedding_info.p", "rb")) # glove word embeddings
train_data, dev_data, test_data = pickle.load(open(path + "/data_info.p", "rb"))  # glove
train_data, dev_data, test_data = pickle.load(open(path + "/data_info_bert.p", "rb")) # bert
train_data, dev_data, test_data = pickle.load(open(path + "/data_info_distilbert.p", "rb")) # distilbert

Models

The code of all models. It is worth noting that 4, 5, 6, 7. The models are directly fine-tuned after naive training, with the support set of testing classes. Each single method can be runed directly

compute_cov_score

The code about how to compute cov score for features.

Citation

@inproceedings{hu-etal-2022-classical,
    title = "Classical Sequence Match Is a Competitive Few-Shot One-Class Learner",
    author = "Hu, Mengting  and
      Gao, Hang  and
      Bai, Yinhao  and
      Liu, Mingming",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.419",
    pages = "4728--4740"
}