Stochastic evolutionary modelling of language change and applications to historical data

Abstract

Language is a complex adaptive system whose properties emerge from a network of communicative interactions between speakers, modulated by their cognitive and social biases. In this conceptualisation, language change has a crucial role as the evolutionary process driving the adaptation of the language system. With the increasing availability of massive digital corpora datasets, stochastic evolutionary models of language change have the potential of providing empirical and quantitative explanations to the nature and origin of human language. In this thesis, I build on and develop such models in order to improve their applicability to historical data and their explanatory potential.

First, I address limitations in the application to historical language data of the Wright-Fisher model from population genetic. By improving on the Beta-with- Spikes approximation to the Wright-Fisher transition probability, I am able to more accurately discern drift from selection in time series of language use. I apply this method to the detection of a phonological bias in the evolution of the past tense of English verbs. I further introduce a methodology able to detect abrupt changes in social dynamics in time series of language use by modelling them as discontinuities in the selective forces acting on the data. I benchmark this method by applying it to the detection of well-documented historical spelling reforms in Spanish.

Secondly, I introduce an Iterated Bayesian Learning model for the evolution of grammatical structure through cultural transmission. Unlike previous evolutionary models of language change, this new model accounts for the coevolution of interrelated linguistic functions. By accounting for nontrivial communicative effects, the model reproduces features of natural languages absent from simpler models of cultural transmission, like communicative effects in the arising stationary distribution of grammars, and broken detailed balance generating directionality in change. I explore directionality by applying entropy production, a concept from non-equilibrium Statistical Mechanics, to a variety of models of language change.

Finally, I show that this model of the cultural transmission of grammatical structure is equivalent to a set of co-evolving and interdependent Wright-Fisher processes, facilitating its application to empirical data. Its application to the evolution of relativisers and to the emergence of the periphrastic use of do in English highlights the potential of evolutionary models as tools for the empirical and quantitative testing of hypotheses in historical linguistics.