Modeling linguistic evolution: a look under the hood

Abstract

This paper takes a detailed look at some popular models of evolution used in contemporary diachronic linguistic research, focusing on the continuous-time Markov model, a particularly popular choice. I provide an exposition of the math underlying the CTM model, seldom discussed in linguistic papers. I show that in some work, a lack of explicit reference to the underlying computation creates some difficulty in interpreting results, particularly in the domain of ancestral state reconstruction. I conclude by adumbrating some ways in which linguists may be able to exploit these models to investigate a suite of factors which may influence diachronic linguistic change.

Exploring horizontal homophony in pronominal paradigms: A case study where cross-linguistic regularities defy individual learning biases

“…Results corroborate a hierarchy of horizontal homophony: 3 > 2 > 1 (where > represents frequency inequality) in the world's languages documented today. Whether or not this hierarchy can also be found diachronically (as we would expect with a robust typological bias) remains to be explored in future work using phylogenetic modelling techniques (Cathcart, 2018;Greenhill et al, 2020;Bickel, 2015;Jäger & Wahle, 2021).…”

The evolution of similarity avoidance: a phylogenetic approach to phonotactic change

“…Standard approaches for dealing with this problem include stratified sampling, limiting analyses to one language per genetic grouping (Dryer, 2000) and mixed-effects regression, controlling for historical and spatial relatedness when estimating crosslinguistic preferences for a feature or the effect of one feature on another (Jaeger et al, 2011;Naranjo & Becker, 2022). These approaches can be contrasted with phylogenetic models that explicitly specify the transmission and diffusion processes thought to give rise to the diversity we observe, rather than treating them as nuisance factors (Cathcart, 2018).…”

Language follows a distinct mode of extra-genomic evolution

“…A famous case is known as "Jespersen's cycle", where negation oscillates between short and long forms: English, for example, started with "ne", extended it to "na wiht" and similar expressions meaning "not a thing", then shortened it again to "not" and "n't", and now colloquially expands it again by "nothing" in some dialects (as in "she didn't do nothing"). In modern work, this cyclic dynamic is successfully modeled by (first-order) Markov processes [48][49][50][51][52] that are ergodic [48], i.e., all states in a transition chain can be reached in finite but randomly varying time (Figure 1A). This means that there can always be languages in all states (Figure 1B).…”