Explaining Color Term Typology as the Product of Cultural Evolution using a Bayesian Multi-agent Model Mike Dowman ([email protected]) School of Information Technologies, F09, University of Sydney, NSW2006 AUSTRALIA Abstract Basic color terms have prototype properties (Taylor, 1989) as there is usually a single color, the prototype, which An expression-induction model was used to simulate the speakers of the language consider to be the best example of evolution of basic color terms in order to test Berlin and the color term. Colors become increasingly less good Kay’s (1969) hypothesis that the typological patterns examples of the color category as they become more observed in basic color term systems are produced by a dissimilar to the prototype, and the category boundaries are process of cultural evolution under the influence of universal fuzzy, as speakers are unsure about the exact range of colors aspects of human neurophysiology. Ten agents were simulated, each of which could learn color term denotations denoted by each color term. by generalizing from examples using Bayesian inference. There has been a considerable amount of research into the Conversations between these agents, in which agents would properties of basic color terms cross-linguistically. Perhaps learn from one-another, were simulated over several the most important study was that of Berlin and Kay (1969). generations, and the languages emerging at the end of each They examined a sample of 98 languages, and found that simulation were investigated. The proportion of color terms there was very wide variation between the color terms of of each type correlated closely with the equivalent different languages, in that the actual ranges of color frequencies found in the world color survey, and most of the denoted by each term differed between languages. However, emergent languages could be placed on one of the they found that this variation was certainly not completely evolutionary trajectories proposed by Kay and Maffi (1999). random. While the number of color terms varied between The simulation therefore demonstrates how typological patterns can emerge as a result of learning biases acting over languages, which combination of color terms existed in any a period of time. given language seemed to be at least partly predictable. Berlin and Kay found that all languages have between 2 Introduction and 11 basic color terms. For 20 of the languages in their study, they asked informants to map both the outer This paper describes computational modeling experiments boundary of each of the basic color terms on an array of performed to explain the typological patterns observed in Munsell color chips, and to identify the best or most typical the color term systems of human languages. Color terms are example (the prototype) of each term. They discovered that simply words which are used to denote the property of the boundaries of the areas of color denoted by color terms color, and in most languages, a special subset of such words varied greatly between languages, but that the locations of can be identified, which Berlin and Kay (1969) named basic the prototypes of most basic color terms were clustered in a color terms. Berlin and Kay listed a number of criteria few parts of the color space. which they used to distinguish basic color terms from other A further finding emerged when Berlin and Kay words used to denote color. They considered color terms to investigated the combination of color terms existing in any be basic only if they were known by all speakers of the particular language. They produced the implicational language and were highly salient psychologically, and if hierarchy shown in Figure 1 to explain the regularities they did not just name a subset of the colors denoted by which they found. All languages appeared to have terms another color word and their meanings were not predictable with their prototypes at black and white, but some languages from the meanings of their component parts. Application of had no other basic color terms. However, if a language had a these criteria seems to distinguish clearly between basic and term for any of the colors further right in the hierarchy, it non-basic color terms in most languages, although there can always had terms for all the colors appearing to the left of still remain some questionable cases. English has 11 basic that point. color words, red, yellow, green, blue, orange, purple, pink, Berlin and Kay proposed that this hierarchy described the brown, grey, black and white. Terms such as crimson, general patterns seen in color term systems cross- blonde and royal blue are not considered to be basic. linguistically, but they did acknowledge the existence of purple white green pink red black yellow blue brown orange grey Figure 1. Berlin and Kay’s Implicational Hierarchy. 360 some exceptional languages which could not be placed on However, this proposal appears to be too restrictive, because the hierarchy. Since Berlin and Kay published their original it implies a limited set of universal color categories, while study, there has been a great deal of interest in basic color the empirical evidence shows that the boundaries of color terms, and much more data has been collected. These term denotations are quite variable, and that it is only the studies have in large part confirmed Berlin and Kay’s prototypes of color terms that are consistent across original findings, though several modifications have been languages. made to the hierarchy to accommodate languages of types Four special red, yellow, green and blue hues have also which were not attested in the original study. been identified by psychological evidence, and they are A very large survey of the color term systems of 110 termed unique hues, because they can’t be described as minor languages, the World Color Survey (Kay, Berlin, and blends of other colors. Heider (1971) showed that children Merrifield, 1991), has now produced a wealth of high are more likely to pick unique hues rather than other colors quality data allowing a much more complete picture of color out of selections of color chips, and so she proposed that term systems worldwide to be obtained. Using this new unique hues are especially salient. In Heider (1972) she data, Kay and Maffi (1999) produced a new classification of showed that people are best able to pick out a previously color term systems, which has modified the original seen color from an array of color chips when that color is a hierarchy of Berlin and Kay (1969) considerably, but which unique hue, which suggested that unique hues can be still shows that the attested color term systems are only a remembered more easily than other colors. We should note small subset of those which are logically possible. that Heider did not distinguish between the four unique hues There appear to be six fundamental colors, corresponding and the colors which formed the prototypes of color terms to the colors which are usually the prototypes of red, yellow, such as purple and orange, but a large amount of other green, blue, black and white color terms. The order of evidence points to the special status of unique hues. Also emergence of terms containing these fundamental colors is Kay and Maffi (1999) have questioned whether the unique fairly predictable, but the order of emergence of other basic hues apparent from the linguistic and psychological studies color terms, such as purple and orange terms, is less really correspond to those found using neurophysiological predictable. Kay and Maffi’s (1999) classification of color techniques. However, the consensus of opinion seems to be term systems was made by considering only terms whose that the four unique hues have a special psychological denotation included at least one of the fundamental colors, status, regardless of its cause. but Kay, Berlin and Merrifield (1991) noted that purple and The computer model described in this paper is a kind of brown terms tend to emerge before orange or pink ones. expression-induction model. These models, the first of Kay and Maffi (1999) found that 83% of the languages in which was described in Hurford (1987), aim to simulate the World Color Survey lie somewhere along the trajectory cultural evolution of language, usually over a number of shown in Figure 2, which represents a progression in which generations. They contain several agents, each of which is languages evolve from a state in which they contain only capable both of learning some aspect of language, and also two basic color terms, to states in which each of the using the language which they have learned to express fundamental colors is represented by a different basic color themselves, hence creating example utterances from which term. Kay and Maffi also proposed side branches to the other agents can learn. Usually expression-induction models main trajectory in order to accommodate some less common are run several times, so that the general properties of the language types, such as those containing yellow-green-blue, languages which emerge in them can be observed. If all the yellow-green or black-blue composites. There were four emergent languages have a particular property which is also languages which Kay and Maffi were unable to place a universal in real languages, or if the emergent languages anywhere on their trajectories, and which were simply show a limited range of variation, reflecting typological classed as exceptions. This paper attempts to address the patterns, then the model can be said to explain why these issue of what causes these typological patterns, by relating universals or typological patterns exist. Expression- them to fundamental properties of the human visual system. induction models have been developed to account for a wide white-red-yellow + black-green-blue range of linguistic phenomenon.