The color communication game

Abstract

There is clear diversity among speakers of a typical language in how colors are named. What is the impact of this diversity on the people's ability to communicate about color? Is there a gap between a person's general understanding of the color terms in their native language and how they understand a particular term that denotes a particular color sample? Seventy English-speaking dyads and 63 Somali-speaking dyads played the Color Communication Game, where the "sender" in each dyad named 30 color samples as they would in any color-naming study, then the "receiver" chose the sample they thought the sender intended to communicate. English speakers played again, under instructions to intentionally communicate color sample identity. Direct comparison of senders' samples and receivers' choices revealed categorical understanding of colors without considering color naming data. Although Somali-speaking senders provided fewer color terms, interpersonal Mutual Information (MI) calculated from color naming data was similarly below optimal for both groups, and English-speaking dyads' MI did not improve with experience. Both groups revealed superior understanding of color terms because receivers showed better exactly-correct selection performance than was predicted by simulation from their senders' color-naming data. This study highlights limitations on information-theoretic analyses of color naming data.

Figure 1

Methods and stimuli. (A) the Color Communication Game. (B) the colors used in this study, shown with the corresponding Munsell color samples from the World Color Survey. (C) The sender (a Somali speaker in this example) views a sample from his source set of samples and sends a message (either a single color term, or “I don’t know”) to the receiver. (D) The receiver hears or sees the term then selects, from her destination set of color samples, the one she thinks the sender intended to communicate. Neither player can see the other’s samples (shown by a baffle in A). The message improves the receiver’s likelihood of selecting the correct sample, but it does not guarantee success. See Sect. “Experimental methods” and Supplement for further details.

Figure 2

Comparison of color naming results for English (naïve: blue, experienced: green) and Somali (red) color terms, as functions of color sample, sorted in order of hue, then lightness. Color terms below abscissas in C and D correspond to English basic color terms. Black dashed lines highlight samples named with high consensus in both English and Somali. Red dashed lines indicate colors named with high consensus in English but not Somali. (A, C) consensus: the fraction of players who used the most common term for each sample; consensus varied by hue (F_1,86 = 7.649 p = 0.007; hue accounted for 8.2% of the variance in consensus), but the average consensus did not differ across data sets (F_2,86 = 0.888, NS; data sets accounted for 2% of the variance in average consensus). (B, D) number of distinct terms per sample, averaged across players; the number of terms per sample varied by hue (F_1,86 = 4.049, p = 0.047; hue accounted for 4.2% of the variance in term number) and across data sets (F_2,86 = 4.097, p = 0.020; data sets accounted for 13.9% of the variance in term number).

Figure 3

Mutual Information, simulations and results, as a function of the number of color terms in the receiver’s vocabulary. Somali-speaking players: red diamonds. English-speaking players: naïve, blue diamonds; experienced, green convex hulls. All data in log₂ units (bits). Columns (A, C, E): self-communication; columns (B, D, F): dyadic communication. First row: Mutual information from color-naming data. Gray lines, theoretical maximum MI from Eq. (4). Second row: Simulated correct performance from color-naming data. Third row: Empirical exactly-correct choice performance. The major diagonal is not a theoretical maximum for exactly-correct performance, because a player could do better by chance.

Figure 4

Accuracy of sample selection. Data points are the receiver’s chosen samples as a function of the sender’s intended samples. Data are color coded by the receiver’s color selections; positions jittered by less than ± 0.5 in two dimensions. Accurate selections, where the receiver chose the sample intended by the sender, fall on the major diagonal. (A–C), selections for self-communication, where the sender terms were provided by the receivers themselves; (D–F), selections for dyad communication, where the terms were provided by a sender other than the receiver. Numbers in the headings are Spearman’s rho, and all are highly statistically significant (p values near zero). Fiducial lines between panels (A–C) join clumps of samples associated the color terms in Fig. 2; arrow shows the chartreuse sample #14 (see Sect. “Discussion”); bars below panels (A–C) indicate groups of samples named by senders that receivers tended to respond to similarly.

Figure 5

Exactly-correct identification performance as a function of predicted performance based on simulations. (A) self-communication performance is not as good as the predicted level of performance for self-communication based on simulations; because the self-communication prediction falls exactly on the major diagonal (Fig. 3C), the data in panels 5A and 3E are identical. (B) dyad communication performance is better than predicted. Lines in A, B are fitted descriptively to the medians of the data, and have no theoretical significance. (C) mean gaps, ± SEM, between the predicted and observed results in (A, B).

Figure 6

Categorically correct choices. Perfect performance is at log₂ for all sizes of the receiver’s vocabulary. The black lines are fitted descriptively to the three data set; they have no theoretical significance. (A) self-communication; mean values were 0.38 bits (Somali), 0.13 bits (English naïve) and 0.13 bits (English experienced) below perfect. (B) communication within dyads; mean values were 1.45 bits (Somali), 0.96 bits (English naïve) and 1.03 bits (English experienced) below perfect.

Figure 7

Samples named “I don’t know” and samples chosen on trials where DK was the message color term in the Somali data set. The distribution of DK messages varied significantly across the sample set (F_29,1410 = 4.747, p < 0.001; the color sample factor accounted for 8.9% of the variance in the samples called DK), but the choices of samples in response to DK were approximately constant (F_{29, 1410} = 0.995, NS; color sample accountedfor 2% of the variance in DK choices). Furthermore, participants’ choices of DK samples were unrelated to their personal use of DK (partial correlation: r = 0.001, p = 0.97, NS).