Automatic Methods for Sound Change Discovery

Abstract

Describing the phonological history of languages has been a central topic in historical linguistics. In this thesis, we develop automatic methods to discover patterns of sound change in different settings of input and output conditions. More specifically, we focus on three challenging tasks: (1) automatic decipherment, (2) automatic decipherment in unsegmented scripts, and (3) automatic sound law induction. We show that a careful model design that implements historical linguists’ priors and intuitions is essential for the success of these methods. In addition, we demonstrate that these computational methods can provide relevant evidence to answer important research questions in historical linguistics.