ASDE: Low-budget text classification via active semi-supervised learning with debiasing training mechanism

Abstract

Semi-supervised learning (SSL) is widely employed in text classification to address the challenges associated with limited labeled data availability. Nevertheless, current SSL methods often exhibit two significant limitations: they typically neglect the crucial process of selecting the initial labeled data effectively, and they fail to adequately mitigate the inherent bias that accumulates due to error propagation during the semi-supervised training phase. To address these shortcomings, we introduce an Active Semi-supervised learning framework with a DEbiasing training mechanism (ASDE). Specifically, ASDE includes a novel task-aware cold-start active data selection component designed to establish a more representative and informative initial labeled set by leveraging task-specific information. Additionally, to combat the detrimental effects of error propagation, we develop a spatial interpolation debiasing mechanism integrated into the training process. Empirical results on four widely used text classification datasets demonstrate the substantial performance gains achieved by our proposed ASDE framework, particularly under low-budget conditions.