Arxiv

LST: Lexicon-Guided Self-Training for Few-Shot Text Classification

Hazel Kim*, Jaeman Son*, Yo-Sub Han

Improves few-shot self-training by guiding pseudo-labels with a refined lexicon, reducing overconfident early errors when only a handful of labeled examples are available.

Overview

Self-training can amplify early mistakes when labels are scarce. LST guides few-shot self-training with a refined lexicon so pseudo-labels are less likely to reinforce overconfident errors.

The approach is aimed at settings with only a handful of labeled examples, where standard self-training often collapses too quickly.

Why it matters

Pseudo-label noise is one of the main barriers to self-training in the few-shot regime. Lexicon guidance offers a lightweight prior that keeps early iterations more stable.

Paper

Read the PDF on arXiv