Morphological Knowledge in Multilingual Large Language Models: A Comparative Analysis of mT5 and ByT5
Keywords
Authors
Issue Date
2024
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
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 probing two pre-trained language models mT5 and ByT5 sharing same model architecture, training objective, and training data -- which only differ in their tokenization strategies: subword tokenization vs. character-level tokenization. Probing the morphological knowledge encoded in these models on 17 languages, our analyses show that multilingual language models learn the morphological systems of some languages better than others, that morphological information is encoded in the middle and late layers, with morphology being present in earlier layers with standard tokenization, yet character-level models eventually yield commensurate morphological knowledge. Finally, we show that languages with more irregularities require a higher proportion in the pre-training data to compensate for the increased complexity.
Description
Citation
Supervisor
Faculty
Faculteit der Letteren