English

Embedding structure matters: Comparing methods to adapt multilingual vocabularies to new languages

Computation and Language 2023-10-30 v2

Abstract

Pre-trained multilingual language models underpin a large portion of modern NLP tools outside of English. A strong baseline for specializing these models for specific languages is Language-Adaptive Pre-Training (LAPT). However, retaining a large cross-lingual vocabulary and embedding matrix comes at considerable excess computational cost during adaptation. In this study, we propose several simple techniques to replace a cross-lingual vocabulary with a compact, language-specific one. Namely, we address strategies for re-initializing the token embedding matrix after vocabulary specialization. We then provide a systematic experimental comparison of our techniques, in addition to the recently-proposed Focus method. We demonstrate that: 1) Embedding-replacement techniques in the monolingual transfer literature are inadequate for adapting multilingual models. 2) Replacing cross-lingual vocabularies with smaller specialized ones provides an efficient method to improve performance in low-resource languages. 3) Simple embedding re-initialization techniques based on script-wise sub-distributions rival techniques such as Focus, which rely on similarity scores obtained from an auxiliary model.

Keywords

Cite

@article{arxiv.2309.04679,
  title  = {Embedding structure matters: Comparing methods to adapt multilingual vocabularies to new languages},
  author = {C. M. Downey and Terra Blevins and Nora Goldfine and Shane Steinert-Threlkeld},
  journal= {arXiv preprint arXiv:2309.04679},
  year   = {2023}
}

Comments

Camera-ready for Proceedings of the 3rd Workshop on Multilingual Representation Learning

R2 v1 2026-06-28T12:16:50.107Z