Research Seminar - June 28, 2002

Modelling the Evolution of Basic Color Terms

Abstract:

Color words have long been of interest to researchers in linguistics and related fields, both because they display prototype properties, and because they conform to semi-predictable typological patterns across languages. The present work implements an explicit computational model of their acquisition and cultural evolution over a period of several generations. The model simulates a community of speakers, each of whom is able to learn the range of colours which can be denoted by a color word given examples of colours to which it has been used to identify. The learning mechanism takes the form of a Bayesian model, and aims to replicate the mechanism which children use to learn color words. The evolution of language over several generations was modelled, by making the simulated people give names to randomly chosen colours, so creating examples from which other agents could learn. Periodically older people would be replaced by new agents who had no knowledge of any color word, and occasionally a speaker would make up a completely new color word so as to allow new words to emerge in the language. The initial results of the simulations reproduced some of the properties of empirically observed color term systems. Firstly, the color words learned by each agent always have prototype properties, with a single best example of the color word, and fuzzy boundaries. Furthermore the languages which emerge almost always partition the color space, so that each color is named by one color term, although there is considerable disagreement between speakers about the exact location of the boundaries between neighbouring color terms. An extension of the model added innate biases based on neurophysiological and psycholinguistic evidence, so that focal red, yellow, green and blue were especially salient, and were more likely to be remembered by speakers as example colours. Simulations under these conditions produced languages which replicated some of the observed typological patterns, with attested types of color term system appearing more frequently than unattested ones. These results provide evidence to support the validity of the evolutionary model, and the Bayesian model of language acquisition on which it is based.