You Can Have Your Data and Balance It Too: Towards Balanced and Efficient Multilingual Models

TL;DR

This paper introduces a teacher-student knowledge distillation method for multilingual models that improves performance on low-resource languages while maintaining high-resource language performance, promoting more balanced NLP systems.

Contribution

The paper proposes a novel multilingual training technique using monolingual teacher models and balanced data to enhance low-resource language performance.

Findings

Outperforms standard training in low-resource languages

Maintains high-resource language performance

Uses the same data amount as standard methods

Abstract

Multilingual models have been widely used for cross-lingual transfer to low-resource languages. However, the performance on these languages is hindered by their underrepresentation in the pretraining data. To alleviate this problem, we propose a novel multilingual training technique based on teacher-student knowledge distillation. In this setting, we utilize monolingual teacher models optimized for their language. We use those teachers along with balanced (sub-sampled) data to distill the teachers' knowledge into a single multilingual student. Our method outperforms standard training methods in low-resource languages and retrains performance on high-resource languages while using the same amount of data. If applied widely, our approach can increase the representation of low-resource languages in NLP systems.