How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models

TL;DR

This paper systematically compares multilingual and monolingual pretrained language models across nine diverse languages and five tasks, revealing the significant impact of tokenizer choice on monolingual performance.

Contribution

It provides a controlled empirical analysis showing the importance of monolingual tokenizers and data size in multilingual model performance, with new insights into tokenizer effects.

Findings

Monolingual tokenizers improve multilingual model performance.

Language representation in vocabulary affects performance gap.

Tokenizer replacement enhances downstream task results.

Abstract

In this work, we provide a systematic and comprehensive empirical comparison of pretrained multilingual language models versus their monolingual counterparts with regard to their monolingual task performance. We study a set of nine typologically diverse languages with readily available pretrained monolingual models on a set of five diverse monolingual downstream tasks. We first aim to establish, via fair and controlled comparisons, if a gap between the multilingual and the corresponding monolingual representation of that language exists, and subsequently investigate the reason for any performance difference. To disentangle conflating factors, we train new monolingual models on the same data, with monolingually and multilingually trained tokenizers. We find that while the pretraining data size is an important factor, a designated monolingual tokenizer plays an equally important role in the downstream performance. Our results show that languages that are adequately represented in the multilingual model's vocabulary exhibit negligible performance decreases over their monolingual counterparts. We further find that replacing the original multilingual tokenizer with the specialized monolingual tokenizer improves the downstream performance of the multilingual model for almost every task and language.