Abstract
Information-theoretic analysis of color-naming data predicts color identification behavior, but it relies on certain assumptions. We examine these assumptions using a “color naming game”, where a “sender” selects a sample from a source set of samples and communicates its identity to a “receiver” by uttering a message color term. The receiver uses the message to identify the sender’s sample. Color communication may be measured using Mutual Information, which depends on two assumptions: (1) the mapping (color samples)<->(color terms) is invariant for each player; (2) the receiver chooses a random sample from those associated with the message color term. Ideally, the number of correct choices equals the number of terms in the receiver’s color vocabulary.----------------Methods/Results----------------Ninety-four Somali-speaking and 31 English-speaking participants actually played the color communication game, as both sender and receiver. We simulated identification performance from the color-naming data, using the two assumptions. Self-communication, where one person was both sender and receiver, was below ideal, but the agreement between prediction and performance was improved by assuming (against #1) that speakers “forgot” a few samples associated with each color term. When the sender and receiver were different people, color identification performance was always better than predicted. Consolidating synonyms within languages by cluster analysis moved the predicted interpersonal communication even closer to ideal, further increasing the gap between prediction and performance. Against #2, both groups of participants chose samples near the centers of their color categories, which further improved predicted Somali speakers’ performance. ----------------Conclusions----------------The “focal colors,” which are “typical colors” located near the centers of the color categories, are important for theories of color naming. The violation of #2, which does not depend on any simulations, suggests that focal terms exist to improve color communication. Information-theoretic analyses using Mutual Information cannot necessarily assume random behavior in the face of uncertainty.