In search of isoglosses: continuous and discrete language embeddings in Slavic historical phonology

This paper investigates the ability of neural network architectures to\neffectively learn diachronic phonological generalizations in a multilingual\nsetting. We employ models using three different types of language embedding\n(dense, sigmoid, and straight-through). We find that the Straight-Through model\noutperforms the other two in terms of accuracy, but the Sigmoid model's\nlanguage embeddings show the strongest agreement with the traditional\nsubgrouping of the Slavic languages. We find that the Straight-Through model\nhas learned coherent, semi-interpretable information about sound change, and\noutline directions for future research.\n