Tokenization and Morphology in Multilingual Language Models: A Comparative Analysis of mT5 and ByT5
This work addresses the problem of optimizing tokenization for multilingual language models, which is incremental as it builds on existing models to analyze morphological encoding.
The study investigated how tokenization affects morphological knowledge in multilingual language models by comparing mT5 (subword tokenization) and ByT5 (character-level tokenization) across four tasks and 17 languages, finding that models learn morphology better for some languages, with morphological information encoded in middle and late layers, and languages with more irregularities benefit from more pre-training data.
Morphology is a crucial factor for multilingual language modeling as it poses direct challenges for tokenization. Here, we seek to understand how tokenization influences the morphological knowledge encoded in multilingual language models. Specifically, we capture the impact of tokenization by contrasting two multilingual language models: mT5 and ByT5. The two models share the same architecture, training objective, and training data and only differ in their tokenization strategies: subword tokenization vs.\@ character-level tokenization. Probing the morphological knowledge encoded in these models on four tasks and 17 languages, our analyses show that the models learn the morphological systems of some languages better than others and that morphological information is encoded in the middle and late layers. Finally, we show that languages with more irregularities benefit more from having a higher share of the pre-training data.