ALP: Data Augmentation Using Lexicalized PCFGs for Few-Shot Text Classification

Authors: Hazel H. Kim, Daecheol Woo, Seong Joon Oh, Jeong-Won Cha, Yo-Sub Han10894-10902

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on few-shot text classification tasks demonstrate that ALP enhances many state-of-the-art classification methods. In this section, we present experimental results on our contributions.
Researcher Affiliation Collaboration 1 Yonsei University, Seoul, Republic of Korea 2 NAVER AI Lab, 3 Changwon National University, Changwon, Republic of Korea
Pseudocode No The paper provides an overview of the ALP algorithm in Figure 2 and describes its stages, but it does not include a formal pseudocode block or algorithm steps formatted like code.
Open Source Code No The paper does not provide an explicit statement or a link indicating that the source code for the described methodology (ALP) is publicly available.
Open Datasets Yes Usual few-shot learning refers to the setup where k samples per class are available for training and other disjoint k samples per class are available for validation, where k is usually small (k-shot learning). Table 2: Comparison of data augmentation methods. We use the Self-Training (ST) semi-supervised learning setup with k-shot samples for both training and validation, where k {5, 10}.
Dataset Splits Yes Usual few-shot learning refers to the setup where k samples per class are available for training and other disjoint k samples per class are available for validation, where k is usually small (k-shot learning). Table 2: Comparison of data augmentation methods. We use the Self-Training (ST) semi-supervised learning setup with k-shot samples for both training and validation, where k {5, 10}.
Hardware Specification No The paper does not provide specific details on the hardware used for experiments, such as GPU/CPU models or memory specifications.
Software Dependencies No The paper mentions software like Fair Seq, BERT-Base classifier, and various SSL approaches (ST, UST, Mix Text), but does not specify their version numbers.
Experiment Setup Yes We use the official code with the recommended insertion, deletion, and swap ratios the authors provided. ... We select German as intermediate languages for back-translation using Fair Seq and set 0.9 as the random sampling temperature. ... We use the default masked proportion and the pre-trained weights provided by the authors. Throughout the experiments we generate 200 samples for all augmentation methods, unless specified differently. We have conducted experiments with 5 random samplings of the labeled data, shuffling of data being presented to the models, and the weight initialization.