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

Tomasz Limisiewicz, Dan Malkin, Gabriel Stanovsky


Abstract
Multilingual models have been widely used for the cross-lingual transfer to low-resource languages. However, the performance on these languages is hindered by their under-representation 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 retains 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.
Anthology ID:
2023.sigtyp-1.1
Volume:
Proceedings of the 5th Workshop on Research in Computational Linguistic Typology and Multilingual NLP
Month:
May
Year:
2023
Address:
Dubrovnik, Croatia
Editors:
Lisa Beinborn, Koustava Goswami, Saliha Muradoğlu, Alexey Sorokin, Ritesh Kumar, Andreas Shcherbakov, Edoardo M. Ponti, Ryan Cotterell, Ekaterina Vylomova
Venue:
SIGTYP
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
1–11
Language:
URL:
https://aclanthology.org/2023.sigtyp-1.1
DOI:
10.18653/v1/2023.sigtyp-1.1
Bibkey:
Cite (ACL):
Tomasz Limisiewicz, Dan Malkin, and Gabriel Stanovsky. 2023. You Can Have Your Data and Balance It Too: Towards Balanced and Efficient Multilingual Models. In Proceedings of the 5th Workshop on Research in Computational Linguistic Typology and Multilingual NLP, pages 1–11, Dubrovnik, Croatia. Association for Computational Linguistics.
Cite (Informal):
You Can Have Your Data and Balance It Too: Towards Balanced and Efficient Multilingual Models (Limisiewicz et al., SIGTYP 2023)
Copy Citation:
PDF:
https://aclanthology.org/2023.sigtyp-1.1.pdf
Video:
 https://aclanthology.org/2023.sigtyp-1.1.mp4