A Formal Model of Phonological Typology

One of the most contested debates in phonology concerns identifying factors that affect typology. Two lines of thought emerge in this discussion: Analytic Bias (AB) and Channel Bias (CB) approach (Moreton, 2008). The AB approach claims that cognitive biases in learning influence the typology, while the CB approach assumes phonetic precursors and transmission of language affect the typology (Moreton, 2008). Both approaches are supported by empirical evidence. Phonological processes that are typologically rare have been shown to be more difficult to learn or their learning requires more input data (e.g. Wilson, 2006; White, 2014, and many others; for a survey, see Moreton & Pater, 2012a,b). On the other hand, processes that result from phonologization of phonetically motivated sound changes with clear phonetic precursors are also typologically more frequent (Blevins, 2004). While an increasing body of work acknowledges that both AB and CB influence the typology (Moreton, 2008), very few attempts have been made to model the two approaches together: research either focuses on one or the other factor (Blevins, 2004; Kiparsky, 2006; Wilson, 2006; White, 2017, i.a.). This paper aims to fill this gap and proposes a model of phonological typology that models both AB and CB influences together. I argue that this unified model performs better in deriving the surface typology and has an advantage over the current “split” models in that it encodes not only that some processes are rare and others more frequent, but also why such typological differences arise. The first step in the direction of the new typological model are quantitative models of the two sub-components: AB and CB. While several attempts have been made to quantify and model learning of phonology (AB; Wilson, 2006; White, 2017), quantitative approaches to CB have received less attention. The most recent attempt in Cathcart (2015) is computationally too demanding and fails to provide implementable quantitative results. In the first part of this paper, I present and expand on a new model of typology within the CB approach that introduces the concept of Historical Probabilities of Alternations (Pχ ) (from Beguš, 2016). In the second part, I propose a model of typology that admits both CB and AB and employs the Maximum entropy probability distribution over candidates (Goldwater & Johnson, 2003) to derive typological probabilities. The new model of typology combines Historical Probabilities of Alternations introduced here with the learnability metric adapted from Wilson (2006) and unifies them in an MaxEnt-compatible framework.