DSML: Dual Semi-Supervised Meta-Learning for Multi-Lingual Text Classification

ICLR 2026 Conference Submission15754 Authors

Keywords: multi-lingual text classification, Meta-learning, Semi-supervised learning

Abstract: Multi-lingual text classification (MLTC) is a challenging task since it faces language differences and scarce annotated data (especially in low-resource language). To improve the model's multilingual comprehension and text classification ability with a small amount of annotated data, this paper proposes a dual semi-supervised meta-learning method (DSML). Specifically, DSML constructs a teacher-student framework and uses dual meta-learning to help the teacher and student collaborative evolution (co-evolution). The teacher and student models are both initialized with text classification ability with a limit annotated data. The teacher model is also initialized with better multi-lingual comprehension ability than the student model. Through a co-evolution mechanism, the student model granularity learns the teacher model's multilingual comprehension ability, ultimately improving both multi-lingual comprehensive ability and text classification ability. We conduct extensive experiments on a newly collected MLTC dataset and the experiments show that our DSML model achieves a state-of-the-art performance. We give a detailed analysis of the experiments and the data/code will be released on GitHub.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 15754