Contrast-Enhanced Semi-supervised Text Classification with Few Labels

Authors: Austin Cheng-Yun Tsai, Sheng-Ya Lin, Li-Chen Fu11394-11402

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform large-scale experiments on five benchmark datasets with merely 30 labeled data for training and validation datasets per class. Table 2: Performance (test accuracy(%)) comparison with baselines.
Researcher Affiliation Academia Department of Computer Science and Information Engineering, National Taiwan University {r08922086, r09944044, lichen}@ntu.edu.tw
Pseudocode No The paper describes its methods using prose and mathematical formulations but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We evaluate CEST on five public datasets (Table 1), including IMDB (Maas et al. 2011), SST-2 (Socher et al. 2013), Elec (Mc Auley and Leskovec 2013) for sentiment analysis and DBpedia (Mendes, Jakob, and Bizer 2012), AG News (Zhang, Zhao, and Le Cun 2015) for topic classification.
Dataset Splits Yes We randomly select 30 labeled data per class with different random seeds for training and validation set and use the test set in original dataset. Table 1: Dataset statistics.
Hardware Specification Yes The results are averaged for three runs, with each run taking 3-8 hours on an NVIDIA RTX3090.
Software Dependencies No The paper mentions using "huggingface's BERT (bert-base-uncased)" but does not specify versions for software dependencies like Python, PyTorch, or the Hugging Face library itself.
Experiment Setup Yes We set the maximum token length in sentences to 256 and clip tokens exceeding the limit. The learning rate is fixed to 1e 5, and hyper-parameters are set to k = 2, τ = 10, λ = 0.75, |SU| = 2000, dim(Z) = 128.