Probing language identity encoded in pre-trained multilingual models: a typological view

Abstract

Pre-trained multilingual models have been extensively used in cross-lingual information processing tasks. Existing work focuses on improving the transferring performance of pre-trained multilingual models but ignores the linguistic properties that models preserve at encoding time—“language identity”. We investigated the capability of state-of-the-art pre-trained multilingual models (mBERT, XLM, XLM-R) to preserve language identity through language typology. We explored model differences and variations in terms of languages, typological features, and internal hidden layers. We found the order of ability in preserving language identity of whole model and each of its hidden layers is: mBERT > XLM-R > XLM. Furthermore, all three models capture morphological, lexical, word order and syntactic features well, but perform poorly on nominal and verbal features. Finally, our results show that the ability of XLM-R and XLM remains stable across layers, but the ability of mBERT fluctuates severely. Our findings summarize the ability of each pre-trained multilingual model and its hidden layer to store language identity and typological features. It provides insights for later researchers in processing cross-lingual information.

Keywords

• Language identity
• Pre-trained model
• Language model
• Typology
• Multilingual model