MCVIE: An Effective Batch-Mode Active Learning for Multi-label Text Classification

Abstract

Data labeling for multi-label text is a challenging task in natural language processing, and active learning has emerged as a promising approach to reduce annotation effort while improving model performance. The primary challenge in multi-label active learning is to develop query strategies that can effectively select the most valuable unlabeled instances for annotation. Batch-mode active learning approaches, which select a batch of informative and diverse instances in each iteration, have been considered useful for improving annotation efficiency. However, challenges such as incomplete information ranking and high computational costs still hinder the progress of batch-mode methods. In this paper, we propose MCVIE, a novel batch-mode active learning method for multi-label text. MCVIE employs a two-stage active learning query strategy. Firstly, we combine two measures of prediction uncertainty and category vector inconsistency to calculate the basic information score for each example-label pair. Then, we use the Euclidean distance of text feature vectors to iteratively select diverse and informative example-label pairs for annotation. Experimental results on three benchmark datasets demonstrate that MCVIE outperforms other competitive methods.