Abstract
Closely related languages show linguistic similarities that allow speakers ofone language to understand speakers of another language without having activelylearned it. Mutual intelligibility varies in degree and is typically tested inpsycholinguistic experiments. To study mutual intelligibility computationally,we propose a computer-assisted method using the Linear Discriminative Learner,a computational model developed to approximate the cognitive processes by whichhumans learn languages, which we expand with multilingual semantic vectors andmultilingual sound classes. We test the model on cognate data from German,Dutch, and English, three closely related Germanic languages. We find that ourmodel's comprehension accuracy depends on 1) the automatic trimming ofinflections and 2) the language pair for which comprehension is tested. Ourmultilingual modelling approach does not only offer new methodological findingsfor automatic testing of mutual intelligibility across languages but alsoextends the use of Linear Discriminative Learning to multilingual settings.