DDSUD: dynamically detecting subsequence uncertainty and diversity for active learning in imbalanced Chinese sentiment analysis.

Abstract

Sentiment structure analysis in Chinese text typically relies on supervised deep-learning methods for sequence labeling. However, obtaining large-scale labeled datasets is both resource-intensive and time-consuming. To address these challenges, this study proposes Dynamically Detecting Subsequence Uncertainty and Diversity (DDSUD), a Bidirectional Encoder Representations from Transformers (BERT)-based active learning framework designed to tackle subsequence uncertainty and enhance the diversity of imbalanced datasets. DDSUD combines subsequence uncertainty detection, diversity-driven sample selection, and dynamic weighting, enabling an adaptive balance between these factors throughout the active learning iterations. Experimental results show that DDSUD achieves performance close to fully supervised training schemes with only 50% of the data labeled, and outperforms other state-of-the-art active learning methods with the same amount of labeled data. Moreover, by dynamically adjusting the trade-off between subsequence uncertainty and diversity, DDSUD demonstrates strong adaptability and generalization capability in low-resource environments, especially in handling imbalanced datasets, significantly improving the recognition of minority class samples.