Browse Teach Curate

Uncertainty-Guided Self-Training for Few-Shot Toxic Speech Detection

A method (U-GIFT) for detecting toxic online speech when only a handful of labeled examples exist. It couples a Bayesian Neural Network's predictive uncertainty with active-learning sample selection inside a self-training loop: unlabeled items are ranked by uncertainty, the most confident pseudo-labels are added to training, and the classifier is retrained iteratively. The scheme is agnostic to the underlying pre-trained language model and stays robust under class imbalance and cross-domain shift.

KITT Text

U-GIFT: Uncertainty-Guided Firewall for Toxic Speech in Few-Shot Scenario Yihao Wang

U-GIFT is a few-shot toxic speech detection method that reduces reliance on large annotated datasets by combining active learning with Bayesian Neural Networks in a self-training loop. Uncertainty es…