Uncertainty Sentence Sampling by Virtual Adversarial Perturbation

TL;DR

This paper introduces VAPAL, a novel active learning framework that combines uncertainty and diversity sampling using virtual adversarial perturbation, improving data efficiency in sentence understanding tasks.

Contribution

VAPAL is the first method to integrate uncertainty and diversity via virtual adversarial perturbation for active learning in sentence understanding.

Findings

VAPAL outperforms strong baselines on four datasets.

It effectively balances uncertainty and diversity.

VAPAL reduces annotation costs while maintaining high performance.

Abstract

Active learning for sentence understanding attempts to reduce the annotation cost by identifying the most informative examples. Common methods for active learning use either uncertainty or diversity sampling in the pool-based scenario. In this work, to incorporate both predictive uncertainty and sample diversity, we propose Virtual Adversarial Perturbation for Active Learning (VAPAL) , an uncertainty-diversity combination framework, using virtual adversarial perturbation (Miyato et al., 2019) as model uncertainty representation. VAPAL consistently performs equally well or even better than the strong baselines on four sentence understanding datasets: AGNEWS, IMDB, PUBMED, and SST-2, offering a potential option for active learning on sentence understanding tasks.