A feature-based neural model of sound change informed by global lexicostatistical data

Abstract

Historical linguists have successfully reconstructed numerous unattested ancestral languages for over a century, mainly by applying the comparative method, a powerful procedure for recovering extinct languages and understanding how they developed into their modern daughter languages. With the exponential rise of computational power, scholars have been trying to develop computational solutions for tasks in historical linguistics for roughly two decades. The success of these methods, however, is limited to solving some individual tasks satisfyingly, while there are still no good solutions for other tasks. Part of the reason why scholars were not able to find good computational methods for some parts of the comparative method is that historical linguists often rely on their intuition and general linguistic knowledge when reconstructing ancestral languages, a component that computational models naturally lack. This thesis presents a neural model that aims at bridging that gap by providing typological information about the likelihood of sound changes. The model was trained on large-scale global lexical data and is therefore able to assess whether a queried sound change is common or uncommon on a global scale. Since it operates on phonological features, it is able to process any given sound change between two arbitrary IPA symbols. The model was trained on sound changes observed in Maximum Parsimony reconstructions on a large-scale global lexical dataset. The model was trained as a binary classifier in a noise-contrastive estimation setting, where the observed sound changes contributed positive training data which was weighed against randomly generated negative training data. Applying a weighted version of Maximum Parsimony, in which the weights were derived from the model, produced better reconstructions for Proto-Austronesian and Proto-Oceanic than unweighted Maximum Parsimony reconstructions. That showed that the model was able to learn common sound changes, including the direction in which they tend to happen. While it requires further systematic testing, the model shows the potential to enhance tasks in computational historical linguistics by simulating implicit linguistic knowledge as a component of the comparative method.