Get started
Work

Mutual Intelligibility and Receptive Multilingualism in Large Language Models: A Demonstration Through Slavic Languages

Year: 2026
Type: article
Abstract: This thesis concerns itself with the usage of Mutual Intelligibility as a metric of language similarity in the context of Large Language Models. The purpose of this is to explore potential data augmentation and training methods related to Low Resource Languages and Extremely Low Resource Languages, languages which do not have sufficient data for modern Natural Language Processing methods such as Large Language Models. Such languages are considered at-risk within the NLP world due to a variety of factors beyond the obvious difficulty in training for them, such as Model Collapse and Cultural Homogenization. Using Slavic languages as an example, this thesis explores if Mutual Intelligibility within humans is accurately represented by LLMs by replacing human participants with LLMs in an otherwise identical experiment to measure the Mutual Intelligibility of languages. A selection of off-the-shelf GPT and BERT models trained in Croatian, Czech, Polish, Slovak, and Slovene are tested using a series of Cloze tests in each of the other languages. The results are compared to results produced by the work done with humans by Gooskens et al., from which the Cloze tests are taken, through means of descriptive analysis, cosine similarity metrics, and statistical correlation. It is shown that Mutual Intelligibility as found in the tested LLMs does not necessarily represent the state of human language, but there is some resemblance and correlation. Thus, it should still be considered as an important similarity metric for consideration in methods such as cross-lingual transfer training. Additionally, this thesis covers a major incongruence between human capabilities and that of LLMs by exploring how tokenization is affected by the Orthographic Intelligibility of languages, and that using a simple transliteration based pre-processing step can reduce the discrepancy in results.
Language: English
Open Access status: green
OpenAlex ID: W7208801461
FWCI:
Cites: 28
Cited by:
Related to: 10