Statistical physics of language change inferred from time evolving maps

Abstract

Language change may be studied by tracking linguistic variant frequencies through space and time. Observed frequencies may be viewed as realisations of a statistical field. We first set out an efficient cross validated method for inferring the empirical field --- the variant frequency distribution estimated directly from data --- in our case, high resolution surveys from 20th century USA. We then derive a model of the statistical field, assuming that individuals diffuse locally, periodically migrate, and copy one another based on a form of replicator dynamics involving both bias and linguistic accommodation. We infer model parameters by comparing changes in model and empirical fields. Our inferences demonstrate that spatial linguistic patterns can be rapidly destroyed by migration but that accommodation can create them, protect them, and in combination with diffusion, produce partially predictable pattern dynamics. The results provide the strongest evidence yet for surface tension driven phase ordering in human dialects. We quantify the model's forecasting ability up to a time horizon of twenty five years using the Kullback-Leibler divergence between forecasts and observations.