The color lexicon of the Somali language

Abstract

This empirical study had three goals: (a) to describe Somali color naming and its motifs, (b) to relate color naming by Somali informants to their color vision, and (c) to search for historical and demographic clues about the diversity of Somali color naming. Somali-speaking informants from Columbus, Ohio provided monolexemic color terms for 83 or 145 World Color Survey (WCS) color samples. Proximity analysis reduced the 103 color terms to the eight chromatic color meanings from the WCS plus black, white, and gray. Informants' data sets were grouped by spectral clustering analysis into four WCS color naming motifs named after the terms for the cool colors: (a) Green-Blue, (b) Grue (a single term meaning “green or blue”), (c) Gray, and (d) Dark. The results show that, first, the Somali language has about four motifs among its speakers. Second, individuals' color vision test results and their motifs were not correlated, suggesting that multiple motifs do not arise from individual variation in color vision. Last, the Somali color lexicon has changed over the past century. New color terms often came from the names of familiar colored objects, and informants' motifs were closely related to their ages and genders, suggesting that the diversity of color naming across speakers of Somali probably results from ongoing language change.

Introduction

People around the world differ greatly in how they name colors. Most people who live in highly industrialized societies use a limited lexicon of about 11 basic color terms (Berlin & Kay, 1969) plus four to nine additional color terms (Boynton, 1997; Lindsey & Brown, 2014). In contrast, many people who live in less industrialized societies use color lexicons containing fewer color terms. The universalists (e.g., Berlin & Kay, 1969; Kay, Berlin, Maffi, Merrifield, & Cook, 2010; Regier, Kemp, & Kay, 2015) hold that the color terms in a language's lexicon are drawn on a limited lexicon of universal color terms, which probably are ultimately traceable to the physiology of the visual system. The linguistic relativists hold that the color terms are largely free to vary across languages (Davidoff, Davies, & Roberson, 1999; Lucy, 1997; Lucy & Shweder, 1979; Roberson, 2005) but are constrained by the continuity of color space and may originate from the colors of items in the cultural or natural environment (Levinson, 2000). A related issue is whether the color terms in a language partition color space exhaustively, allowing a person to name, with a single color term, every color he or she might encounter (Kay et al., 2010; Levinson, 2000; Lindsey, Brown, Brainard, & Apicella, 2015). According to most investigators (e.g., Kay & Maffi, 1999; Wierzbicka, 2008), the number of color terms in a language's lexicon is related to the technological or cultural needs of the people who speak the language. No single view—universalist or relativist—can account for all the differences among cultures in their color lexicons. 

Worldwide diversity of color naming

The World Color Survey (WCS; Kay et al., 2010) is a corpus of color naming data provided by about 24 informants speaking each of 110 unwritten world languages. WCS informants showed great diversity in how they named colors, yet there were only about 11 distinct color categories in the WCS data set (Lindsey & Brown, 2006; Regier, Kay, & Cook, 2005), in good general agreement with the universalists' claims. Kay and colleagues (Berlin & Kay, 1969; Kay et al., 2010) proposed that these color categories occur together lawfully in about seven or eight distinct, universal color naming systems. In a similar vein, Lindsey and Brown (2009) showed that their WCS color categories occurred in only about four distinct color naming systems (referred to as motifs; Figure 1), which differed mainly in how the cool colors are named. The four motifs were a Green-Blue motif (in which informants had distinct terms for green and blue), a Grue motif (a single term meaning “green or blue” instead of separate terms for green and blue), a Gray motif (in which the term used for cool-colored samples also named one or more of the gray samples), and a Dark motif (in which the cool-colored samples were called by the term for black).¹ These four motifs recurred, with minor variation, across speakers on all continents. 

Figure 1

 

The WCS color samples and the four motifs. (A) The Munsell color samples shown in the chart used throughout this paper. (B–E) The four motifs of the WCS, shown as concordance maps, for stimuli whose true Munsell hues, chromas, and values of the stimuli (small rectangles) are keyed to the samples in the corresponding positions in panel A. The false colors in panels B through E correspond to the majority color terms, and the lightnesses indicate the fraction of informants, within each motif, who used the majority color terms for the corresponding color samples. Here, gray is false-colored chartreuse and black is false-colored orange.

Many investigators have considered how languages differ in the number of basic color terms in their lexicons (e.g., Berlin & Kay, 1969; Kay et al., 2010; Regier, Kay, & Khetarpal, 2007; Roberson, Davies, & Davidoff, 2000), a view that might suggest that some languages use the Green-Blue motif, whereas others use the Grue, Gray, or Dark motifs. Contrary to this view, there is mounting evidence that there was considerable diversity among speakers within most WCS languages (Bimler, 2007; Lindsey & Brown, 2009; Lindsey et al., 2015; Webster & Kay, 2005, 2007) and many other world languages (Davies & Corbett, 1994; Davies et al., 1992; Davies, Roling, Corbett, Xoagub, & Xoagub, 1997; Heider, 1972; Levinson, 2000; Özgen & Davies, 1998). According to Lindsey and Brown (2009), most of the diversity across languages came from variation in the numbers of informants using each of the distinct motifs. 

Lindsey and Brown's analysis of the WCS was retrospective: The data had already been collected, and they examined the data to determine what structure might exist in the data set. For any retrospective analysis, there is always the possibility that the result is a statistical artifact due to capitalization on chance. Conclusions are always more convincing if the results they depend on can be replicated prospectively. Therefore, the first objective of the present study was to collect prospective data on Somali, which is an African language, and to determine whether it shows evidence of multiple motifs among its speakers. 

The origin of worldwide color naming diversity

Beginning with Gladstone (1858), many authors (reviewed in A. M. Brown, 2015) have speculated that at least part of the worldwide variation in the naming of colors might be due to variation in color vision across people. More recent investigators have suggested that color naming variation among people living near the equator, particularly in the use of words that mean blue, might be related to an acquired Type III (blue type) color vision defect (Bornstein, 1973; Lindsey & Brown, 2002; Ratliff, 1976). Therefore, the second objective of this study was to look for an association between color vision and color naming in a group of African participants. 

The Type III color vision defect is defined by blue-type errors on the Hardy, Rand, Rittler (HRR) pseudoisochromatic plates and the D-15 and FM-100 arrangement tests. These errors are caused by some combination of two phototoxic effects of the ultraviolet (UV) B energy in sunlight on the eye: The ocular lens turns yellow and eventually brown (“brunescent”), blocking short-wavelength light from reaching the retina (Birch, Chilsolm, Kinnear, Marré, et al., 1979), and the short-wavelength-sensitive (S) cones are damaged. The yellowing of the lens can cause blue things to look dark, greenish, or brownish, and damage to the S-cones can lead to tritan-type losses in color discrimination. Lindsey and Brown (2002) reported colorimetric and experimental evidence suggesting that high optical density of the ocular lens could reduce the number of samples that are called blue. Subsequently, other investigators (Hardy, Frederick, Kay, & Werner, 2005) examined the naming of colors among elderly Californians, who had moderately brunescent lenses, and found no evidence of a reduced range of colors called blue. However, California is not on the equator, and few people in California are outdoors for most of their waking hours, so the possibility remains that a heavier dose of UV B might produce denser lenses and color vision impairment that is sufficient to affect the naming of colors. 

The controversy surrounding the Type III color vision deficiency hypothesis is most likely to be resolved by examining the color vision and color naming of people who grew up near the equator, where the daily dose of UV B is the highest in the world. Here, we compare the color naming data and the clinical color vision test results obtained on the same Somali participants. Although there is no way of disentangling the two components of the Type III defect using standard clinical tests, it is easy to determine whether blue-type errors on the color vision tests are common and, if so, whether they are age related, as might be expected of an acquired defect. If they are common, the question then arises whether there is a statistical association between blue-type color vision test errors and the color terms used to name the cool colors, as the color vision hypothesis predicts. 

Language change and color naming

Another possible explanation for the origins of the diversity in color naming within a language community is that the language is changing more generally and that the change in the color lexicon is just one example of this change. The change in Somali color naming cannot be traced using empirical data, as the present study is the first empirical look at color naming in the Somali language. However, there are two good historical dictionaries available (Abraham, 1964; de Larajasse, 1897) that can provide historical linguistic clues about how Somali color naming has changed since the turn of the 20th century. Classic work in sociolinguistics has shown that language change in American and Canadian English follows a particular sociolinguistic trajectory, with new pronunciations, morphosyntactic forms, and vocabulary appearing first in the speech of young women (Labov, 1990; Tagliamonte & D'Arcy, 2009), although evidence on this point is less firmly established in non-European languages (but see Haeri, 1995). Therefore, the third objective of this study was to look for historical evidence and possible effects of demographic variables (age, gender, occupation, and geography) that might be related to the diversity of the color lexicon across speakers of the Somali language. 

Overview of this study

The Somali language is in the Cushitic language family. It is spoken on the horn of Africa and in Somalia, Djibouti, the border region of Kenya, and the neighboring disputed region of Ethiopia (range of latitude: approximately 2° S to approximately 12° N; range of longitude: approximately 38.2° E to 51.25° E; see Lewis, Simons, & Fennig, 2015) Since 1991, civil life in Somalia has been dangerous, and many Somali people immigrated to the United States and Europe as refugees. Columbus, Ohio, is home to about 50,000 immigrants of Somali origin, many of whom have not yet learned to speak fluent English. Here we report on their color vision and color naming, as evaluated in a laboratory setting. We collected color naming data using methods that paralleled those of the WCS as closely as possible while maintaining good experimental control over the instructions, the lighting of the samples, and other details of the procedure. We also tested our informants' color vision using standard clinical procedures, and we collected basic demographic information. 

Previous work on the Somali color lexicon

Somali is one of the 98 languages that Berlin and Kay examined in their 1969 monograph, using work by de Larajasse (1897) and Kirk (1905). They list Somali terms for black, white, red, and green, placing Somali at Stage IIIA in their taxonomy. Somali was only commonly written for about 19 years: between 1972, when modern writing using the Latin alphabet was introduced, and 1991, when the Somali civil war began. Over that brief period, public education flourished and there was a proliferation of books and periodical literature. Maffi (1990) published a scholarly paper on Somali color terms near the end of that period. She concluded that the Somali language had terms for the six Hering primaries (white, black, red, green, yellow, and blue), with two words for yellow and regional variation in the term for red. Based on interviews with Somali-speaking collaborators, consistency across sources, philological evidence, and historical linguistic principles, she concluded that white, black, and red were particularly fundamental and had been in the Somali language for a long time, whereas terms for other colors were added more recently. Because of the ongoing upheaval in Somalia, fewer printed materials in the Somali language have appeared in the past 25 years, although there is an active presence of written Somali on the Internet. 

Experiment 1: Color naming

To achieve the first goal of this study, we collected a set of color naming data provided verbally by Somali-speaking informants, who named physical color samples in the laboratory. Data were collected between February 2011 and April 2013. 

Method

Participants

Informants were recruited from the Somali community of Columbus, Ohio. Coauthor AI interpreted for the experimental sessions and provided linguistic advice on the meanings of Somali color terms. There were two phases to the study: a pilot phase involving 15 informants and 83 color samples, and a main data set involving 26 informants and 145 color samples. 

Informants were between 24 and 86 years old (median age = 56 years); there were 15 females (average age = 45.67 years; SD = 14.42) and 28 males (average age = 58.59 years; SD = 18.52; Figure 2A). Informants 7 and 29 (both males) were diagnosed with red-green color vision deficiency, and their data were eliminated from this report (Appendix C), leaving a total of 26 males in the data set of 41 informants. Informants had lived in the United States for less than 12 years (median = 7 years; Figure 2B), generally after a stay of several months to many years in Kenya. Their places of residence in Somalia (see Figure 2C) were divided by k-means analysis into four geographic regions for statistical analysis. When in Somalia, some informants were nomads who herded camels and goats, some were farmers, some were businessmen, and some worked in the home as students or housewives (Figure 2D). Almost all reported being outdoors most of their waking hours (Appendix D). One informant (#2p) said that he had cataracts, which were due to be removed 3 days after testing. His results are discussed in Experiment 2. 

Figure 2

 

Demographic data. (A) Distribution of the ages of men and women. (B) Distribution of the number of years men and women lived in the United States. (C) Informants' hometowns, divided into four regions for statistical analysis. (D) Informants' occupations in Somalia before immigration to the United States.

All informants were native speakers of the standard dialect of the Somali language. A few informants knew a few words of modern spoken Arabic or Italian, but none claimed conversational knowledge of these languages, and none spoke any Swahili or French. Only two informants had any significant knowledge of English. One of these was AI, the interpreter and coauthor on this article. He was not aware of the purpose of the study at the time of his test, and his data are included in all analyses. None of the other informants except these two had even conversational knowledge of English. Their informed consent and participation in the study depended on coauthor and interpreter AI. 

Materials

The samples used in these experiments were from the Munsell Book of color, Glossy Edition (Munsell Color Services, Gretag Macbeth, LLC, New Windsor, NY). There were two sets of Munsell color samples in the color naming experiment. Both were subsets of the colors in the WCS stimulus set (listed in Appendix A). The pilot data (15 informants: AI and informants #2p–#15p) were collected using a set of 83 samples from two rows of 40 Munsell hues at values 6/ and 8/ and high chroma, plus single black, white, and gray samples (Figure 3A). After the pilot experiment, we expanded our data set to cover more of the WCS stimulus set, particularly to include good examples of red and brown, which were not included in the pilot stimulus set. The main data set (27 informants: informants #1–#6 and #8–#28, excluding color-deficient informants #7–#29) was collected using a set of 145 samples: 20 Munsell hues in each row, at every-other-sample spacing, for seven rows of the WCS stimulus set, with alternate rows staggered in such a way that the sampling was also every other row (Figure 3B). The samples were mounted in gray paper holders (Gray 4.5, Color-aid, Hudson Falls, NY; Figure 3C) and were presented on a tabletop covered in the same gray paper. All samples were illuminated at 2400 lx by four T7 full-spectrum fluorescent lamps (F32T8950, Philips, Amsterdam, the Netherlands; Correlated Color Temperature = 5000 degrees Kelvin; Color Rendering Index = 98; Exposure Value EV₁₀₀ = 9.91). 

Figure 3

 

The color samples used in this study. (A) Colors from the pilot data set and (B) colors from the main data set, shown in their traditional locations within the WCS stimulus diagram (see Figure 1A). (C) Examples of the stimuli, as presented to the informants.

Some investigators have criticized the use of Munsell samples in color naming research because the colors in the Munsell stimulus set are more saturated than those generally found in nature (Lucy & Shweder, 1979). Our choice of samples was dictated by two concerns. First, we wanted our data to be as comparable as possible to previous color naming data sets, and our choice of color samples is similar to those used in a long line of color cognition research, including R. W. Brown and Lenneberg (1954), Berlin and Kay (1969), and the World Color Survey (Kay et al., 2010). Second, we wanted to make sure we found all the common color terms in Somali. Other research using the Optical Society of America (OSA) color samples, which were less chromatically saturated (i.e., lower chroma), failed to turn up any color categories not revealed by work carried out using the Munsell samples we used here (compare Boynton [1997] with Lindsey & Brown [2014]). 

It has also been claimed that the particular choices of chroma in the WCS sample set bias subjects' choices of focal colors (i.e., best examples of colors included in a particular subject's named color categories). Our view is that focal color selection concerns theories of category formation (e.g., Heider, 1972) rather than color naming. Moreover, one group of investigators has shown recently that a chroma model fared poorly in accounting for focal color selection in the WCS (Abbott, Regier, & Griffiths, 2012). 

Procedures

This research was performed under a protocol approved by the institutional review board of The Ohio State University, and it adhered to the tenets of the Declaration of Helsinki. Each informant provided informed consent using forms that were printed in Somali; AI read the forms and explained them to the informant in cases where the informant was not a skilled reader. Each informant provided basic demographic data: age, gender, place of residence in Somalia, dialect of Somali, occupation while in Somalia, the amount of time spent outdoor in Somalia, and the number of years in the United States. 

We presented the color samples, one at a time, in a fixed pseudorandom order (see Tables A1 and A2). We used this type of order because (a) it was used in the WCS, (b) it made the mechanics of presenting the paper samples and keeping track of the responses much easier, and (c) it prevented the cognitive order effects that might occur if a systematic order had been used. The instructions were to name each color sample with a word in the Somali language. That word should be a single standard color term that participants would use to denote the color when speaking to, for example, a brother or sister. It should not be the name of “something,” and the informant should avoid animal colors. 

Results

Pooling across the pilot and final data sets, informants used an inventory of 103 single color terms, 48 (47%) of which were used by two or more informants (63 terms in the pilot data set, with 42% used by two or more informants; 76 terms in the final data set, with 51% used by two or more informants; 36 terms in both data sets). A selection of the frequent terms is listed in Table 1. The modern meanings were provided by coauthor AI, a professional interpreter who is a native speaker of Somali and perfectly fluent in U.S. English. The historical data are from de Larajasse (1897) and Abraham (1964), which are the most comprehensive dictionaries available for their respective eras. 

Table 1

 

Modern and historical translations of Somali color terms.

 

In agreement with Berlin and Kay (1969) and Maffi (1990), several informants volunteered that classic Somali had terms only for black, white, and red. In the present data set, these were always native terms (Table 1; Figure 4). Many of the other modern Somali color terms from our data set were also present in the older dictionaries, but generally they were not listed as color terms. Instead, they were often associated with other culturally or ecologically significant items in the environment, much as Levinson (2000) suggested. For example, our color naming data set includes eight words for yellow. Of the terms generally accepted as native Somali, cawl² meant a species of gazelle in de Larajasse (1897; Maffi, 1990) but was the term for yellow by 1964 (Abraham, 1964), dambar meant colostrum, and ubah meant a fruit-tree flower. Huruud is listed historically as turmeric or saffron but was likely originally a loanword from Arabic or Persian (Cardona, 1988), and sacfarran is clearly related to words for saffron in many languages. 

Figure 4

 

Distribution of color term usage in the main data set, shown within the WCS color chart (see Figure 1A). Color terms are grouped according to their most usual modern color meaning. The lightness of the small rectangles (after 0.65-power compression for better visibility of low-frequency usage) codes the number of informants who named the corresponding sample, with the Somali color term indicated by the label and surrounding color key.

Dahabi is from dahab, which means the metal gold in Somali and Arabic. The two most common terms for yellow in the present data set, jaale and yalow, are recent loanwords (from Italian giallo and English yellow, respectively). Our data set lists four main words for green: cagaar, doog, naq, and aqdar. Cagaar, doog, and naq appear in de Larajasse (1897) and Abraham (1964) as words for green shoots or grass but not as the abstract color green; aqdar is a loanword from Arabic. 

There was a highly variable level of consensus across informants in the usage of the terms in the present main data set, even for frequently used color terms. All informants in the main data set used madow to name the black sample, but two informants in the main data set also used madow to name significant numbers of the blue samples, including some light blue samples, and one informant in the pilot data set called the black sample dhuzxuli (charcoal in English). Everyone in the main data set used guduud to name red samples, but guduud was also applied to a set of reddish, purplish, or brownish samples, much as the informants did in the WCS (Lindsey & Brown, 2009). Abraham (1964) listed guduud as meaning “brownness.” There were no good examples of red in the pilot stimulus set (Figure 2A), so there the use of guduud was more variable: Four pilot informants did not use guduud at all, and one pilot informant set also used guduud to name a yellow sample and a light green sample. 

In contrast to the high consensus for black and red, there were five words for white that were used by multiple people. The most common was cadaan, the meaning of which did not vary across historical sources: Cadaan has apparently always meant whiteness and is the nominal form of cad, which means “white.” Cadays was used by 10 informants, but only two informants used it to name the white sample. Cadays is better translated today to mean a silvery color (according to AI). Historically it was allied with caday, a bush with whitish bark whose twigs were used as a toothbrush. Ciiro historically meant “sky” and is still used to mean a lightly overcast sky. Three informants in the main data set called the white sample casaan. However, translating casaan as white is problematic because six informants used casaan to name 20 or more samples, including sample G33, a light purple sample; two of those six also named the yellow samples casaan. One informant used casaan to name a range of pinkish and purplish samples. The remaining informants who used casaan used it to name the light samples in the data set. Thus, a simple translation of the word casaan is probably not possible; we list it as pastel in Table 1 and Figure 4. Historically, cas (the adjective) and casaan (the noun) meant “red,” and casaan remains the word for red in the Northern variety of Somali (Maffi, 1990). 

There was also low consensus for blue and green. Most informants used cagaar, which is a native Somali term generally translated into English as green, and buluug, which is a loanword from English that meant “laundry bluing” in Abraham (1964) and is now generally translated as blue. However, the actual usage of cagaar and buluug was highly variable, with considerable overlap between the terms (Figure 5). Eleven Somali-speaking informants used buluug to name one or more samples in the brown part of the diagram. This non-English-like usage of buluug is an example of a general trend for loanwords to be imported into the Somali color lexicon, but often without perfect understanding of what colors the words referred to. A similar result applies to the loanword color terms yalow, garay, and baroon (Figure 4). 

Figure 5

 

The distribution of cagaar (the most common Somali term for green) compared with the distribution for green in English, and the distribution of Somali buluug compared with the distribution of blue in English. The Somali terms cover a larger range of colors than their nominal English translations. All data were compressed with exponent 0.65 to show the range of most frequent and minority uses of each of the terms. The small rectangles (in the Somali data) have been made two samples wide for comparison with the English data, even though only every other sample was presented (Figure 3B).

On the lowest end of the range of consensus, 53% of the terms in the pooled data set were provided by single informants. For comparison, the U.S. English data set of Lindsey and Brown (2014) contains 96 color terms, 35% of which were provided by single informants (after discarding a single extreme outlier; for all the U.S. English data, 51% of 122 terms were used by single informants). These results may be compared with those of Lindsey et al. (2015), where 32% of color terms in English, 49% of color terms in Somali, and 64% of color terms in Hadzane were provided by single informants. 

Glossed color categories

Based on our previous studies of the WCS (Lindsey & Brown, 2006, see appendix 2), we classified (glossed) the Somali color names according to 11 universal color categories (K = 1 . . . 11), eight chromatic (K = 1 . . . 8) and three achromatic (K = 9 . . . 11). 

In the present study, achromatic color terms were glossed directly as white, black, or gray according to whether the color samples received the same color name as the most reflective achromatic (white) sample, the least reflective (black) sample, or any of the nonwhite and nonblack achromatic (gray) samples in the stimulus set. 

In Lindsey and Brown (2006), the WCS chromatic terms were glossed by k-means clustering of the patterns of nonachromatic color samples associated with each color term, t, used by each informant, s, expressed as vectors . In addition to classification of the WCS color terms, an outcome of that cluster analysis was a set of eight chromatic cluster centroid vectors, C_k, each characterizing a different universal color category (for details see Lindsey & Brown, 2006). In the present analysis, each Somali term was classified by determining the WCS centroid, C_k, most similar to , as determined by Pearson correlation. In each case, the Somali term was compared to either 83- or 145-sample versions of the original 330-sample WCS centroids.  

Figure 6 shows the concordance maps for the glossed chromatic categories from the main data set. The main feature of this data set was its close resemblance to the comparable data from the WCS (Lindsey & Brown, 2006, figure 2, column 8). However, unlike the average data from the WCS, the terms for the YELLOW-OR-ORANGE category used by some Somali informants included samples in the purple region, and, conversely, there were scattered instances of terms for the PURPLE and PINK categories used for samples in the yellow and green regions of the color diagram. Eight of the 15 pilot study informants (53%) and 11 of the 26 informants in the main experiment (42%) used single terms to name some of the samples in both the yellow and purple regions of the color diagram. Also, pilot informant 4p used guduud to name a yellow sample and a desaturated greenish sample; for that informant, guduud was glossed to PINK rather than RED. 

Figure 6

 

Color term maps for the glossed color categories corresponding to the eight chromatic color terms in the main data set. The small rectangles indicate the color samples according to their positions in the diagram of Figure 3. The lightness of the rectangles (after normalization to 1.0 for the most frequent cell and power = 0.65 compression for better visibility) codes the relative number of informants who named the corresponding sample, with the Somali glossed color terms indicated by the label and surrounding color key.

Color term popularity

To examine the usage of the color terms in more detail, we created a Zipf-like diagram (Mitzenmacher, 2003), just as we did in our previous work on U.S. English (Lindsey & Brown, 2014). As before, we defined the popularity of a color term to be the number of informants who used it to name at least one of the 83 or 145 color samples (see Table 1 for some popularity values). We sorted the list of color terms in the rank order of their popularity, omitting the terms that were used by single informants and assigning ties to consecutive ranks. In Figure 7, the upper curve shows the logarithm of the popularity of the color terms (circles) as a function of logarithms of their sorted ranks, with the color-coding scheme indicating each term's usual English translation. Previous investigators have shown that, in large language corpora, the Zipf function often has a slope of about −1 for words in general everyday use (Mitzenmacher, 2003). However, in this analysis, we were specifically looking for double-power law behavior because in other language corpora, two exponents often “divide words in two different sets: a kernel lexicon formed by about N versatile words and an unlimited lexicon for specific communication” (Ferrer i Cancho & Solé, 2001). Therefore, we fitted the data in Figure 7 with the following formula, obtaining values for the constants by a least squares criterion:  

 

 

Figure 7

 

Zipf-like diagrams of the popularity of Somali color terms used by two or more informants. The number near each line segment is its slope. Disks (raw, or unglossed, data) fit with a function consisting of two power laws. The color of each disk corresponds to the usual English translation of the corresponding term. Triangles (glossed data) displaced downward by 0.6 log₁₀ unit for clarity fit with a function consisting of three power laws. The color of each triangle corresponds to the glossed color meaning. See text for further details.

The best-fitting constants in Equation 1 described a double-power law function (overall r² = 0.978), with the last term never being the minimum. The first segment (slope = −0.07) was fitted to the five most popular terms, which were madow, cagaar, guduud, cadaan, and buluug (black, green, red, white, and blue, respectively). These five terms are nearly at the 100% popularity required of Berlin and Kay's (1969) basic color terms, and they form the “kernel lexicon” of Ferrer i Cancho and Solé (2001) at N = 5. The second, descending segment corresponds to the “unlimited lexicon” of Ferrer i Cancho and Solé. It had a slope of −1.31, which is similar to our previous results on U.S. English color terms (Berlin & Kay, 1969), where the first two segments showed slopes of 0.0 and −1.2. The steeper slope is closer to the slope of −1.0 that has been reported by others for large language corpora. Consistent with this interpretation, many of the Somali color terms on the steeper limb were ordinary words with other meanings, much as Levinson (2000) suggested. 

The lower curve in Figure 7 (triangles) shows the popularity of the 11 glossed color terms (eight chromatic terms plus three achromatic terms), pooled across the two data sets. When Equation 1 was fitted to the glossed data, the resulting curve required all three segments (r² = 0.998). The first segment was near ceiling (slope = −0.020) for color meanings that were used by nearly all informants and corresponded to the universal categories BLACK, WHITE, and RED. The second segment had shallow slope (−0.27) for five color categories named by most informants: YELLOW, GREEN, BLUE, BROWN, and GRAY. The final steep segment (slope = −2.61) was fitted to three color categories—PINK, PURPLE, and GRUE—which were named by smaller subsets of informants. This general result was similar to the Zipf-like function of glossed color terms in English, which also had a three-segment structure. The near-zero slope of the first three glossed terms (BLACK, WHITE, and RED) suggests that these are particularly well-established color terms, just as many of our informants suggested, and the shallow slope of the second segment suggests that those color meanings are well on their way toward being used and understood by all. 

The color term buluug, which nominally means “blue” in English, was near the beginning of the unglossed data set (circles in Figure 7) because it was used by most informants but with a range of color meanings (Figure 5). Our proximity analysis glossed many instances of buluug to other color categories, moving its glossed category BLUE further down the list (triangles in Figure 7). Four common words for yellow—jaale, huruud, yalow, and dahabi—were combined by the glossing process into a single, more prevalent color category YELLOW, which occurs at higher rank. The difference in slope and order between the glossed color category data and the unglossed data for the corresponding color terms indicates that informants have not yet settled on consensus terms to name many of the color categories found in the WCS. 

Color naming motifs

The first goal of this study was to determine whether the Somali language showed evidence of the multiple motifs we observed in the WCS data set. We used spectral clustering to assign the color naming data from the main Somali data set to the motifs of the WCS (Ng, Jordan, & Weiss, 2002). To prepare the Somali data for this analysis, we filled out each Somali informant's data set to conform to the 330-sample WCS format by duplicating the responses for the 145 named samples into the slots of adjacent unnamed color samples. We then added the Somali data to the WCS data set as a 111th language and performed a global spectral clustering analysis on the combined data set to extract four clusters, or motifs. This analysis recovered the same four color naming motifs as in the WCS analysis of Lindsey and Brown (2009; see Figure 1). For each Somali informant, we then identified the WCS motif that contained his or her color naming data (Figure 8). 

Figure 8

 

Color naming diagrams for 41 informants by motif. The four motifs were identified by spectral cluster analysis of the WCS plus Somali. The layout within each informant's data set was as in Figure 3; the false colors correspond to the universal glosses in Figure 6. Gray is false-colored light gray, and black is false-colored dark gray. Informants are numbered in their order of test. Numbers with “p” indicate pilot study informants; informants with an asterisk (*) spoke U.S. English as well as Somali. Informant AI, the interpreter and a coauthor on this article, was experimentally naïve at the time he was tested.

The results showed that the Somali data could be assigned to the Green-Blue, Grue, Gray, and Dark motifs. The Green-Blue data sets were those that contained distinct terms for the GREEN and BLUE categories and were further separated in Figure 8 into those with and without PURPLE. A distinct Green-Blue-Purple motif appears in Lindsey and Brown (2009) at K = 6. The Grue data sets showed extensive use of GRUE for the cool-colored samples. The three data sets assigned to the Gray motif contained distinct terms for the GREEN and GRAY categories. Informant 19 was in the Dark motif and used madow (“black”) to name a wide range of samples, including some light blue samples. 

To examine the impact of the assumption that if the Somali color terms were grouped into clusters, those clusters would be similar to the motifs of the 110 WCS languages, we repeated the spectral clustering procedure on the Somali data alone. When we assumed three clusters, the data fell into Green-Blue and Grue groups, which were similar to the corresponding motifs revealed in the pooled analysis described above, and a single Neutral group, which united the Somali informants assigned to the Gray and Dark motifs in the main analysis (Figure 9). A Neutral motif also appears at K = 3 of Lindsey and Brown (2009, figure 1). Increasing the number of clusters beyond three yielded smaller groups that differed mainly in the presence or absence of the PURPLE category. It is important to recognize that with only 27 informants in the main data set, the support for any minority clusters we might find would be limited. Nonetheless, our results show that an independent clustering analysis of the Somali data set clearly revealed multiple groups that were similar to the motifs reported by Lindsey and Brown (2009): Green-Blue, Grue, and an achromatic motif we refer to here as Neutral. Moreover, the results of our quantitative analyses (Figure 8) agree visually with what one would obtain on the basis of inspection alone. 

Figure 9

 

Consensus maps of the three motif clusters discovered at K = 3 when the main Somali data set was analyzed in isolation. Other details are as in Figure 1.

Spectral cluster analysis was not appropriate for the pilot data because there is no obvious way to prepare the pilot data to provide names for the whole 330-sample WCS stimulus set. Therefore, we assigned the pilot data to motifs by determining, for each informant's data set, which WCS motif's centroid was “closest” using proximity analysis. This method finds the distance between each informant's data set and each of the motif centroids and then assigns the informant to the motif with the closest centroid (Appendix B). The pilot data appear in the lower parts of Figure 8, in the columns corresponding to their closest motifs. The similarity between the pilot and main data sets is clear by inspection. 

The spectral cluster analysis of the main data set and the proximity analysis of the pilot data relied on different assumptions. To determine how all the Somali motif assignments shown in Figure 8 compare quantitatively to the WCS motifs, we calculated, for each motif group of Somali informants, pooled across both data sets, the average proximity to the centroid of their assigned motif. Then we compared that value to the average proximities of the informants within each motif within each of the 110 languages in the WCS. Somali informants in the Green-Blue motif were in the 52nd percentile of the Green-Blue averages for the WCS languages. Informants in the Grue motif were in the 31st percentile, informants in the Gray motif were in the 39th percentile, and informants in the Dark motif were in the third percentile of the corresponding motif averages within the WCS languages. 

These analyses show that multiple motifs exist side by side among the idiolects of the Somali-speaking community, just as they do in the languages of the WCS. There is strong evidence of a Green-Blue, a Grue, and some form of Neutral motif within the Somali language community, but evidence in favor of a distinct Dark motif is less strong. 

Discussion

The results of Experiment 1 achieved the first goal of this research. The Somali color lexicon has terms that name the three universal color categories (BLACK, WHITE, and RED), which were used by almost all informants. In addition, most informants named about five additional color categories. These categories are well on their way to being named by all, albeit with less-than-perfect consensus about what their color terms should be. A minority of informants named the remaining three color categories. The results of Experiment 1 replicate, in a prospective study on an African language, the finding of Lindsey and Brown (2009) that the color naming systems of world languages are distributed across their speakers according to multiple naming motifs. These motifs are similar worldwide, and the Somali language color lexicon contains at least two—and possibly all four—of the motifs of the WCS. 

Experiment 2: Color vision

The second goal of this study was to compare participants' color naming results to their performance on three standard tests of color vision deficiency. We wanted to test the hypothesis outlined in the Introduction that age-related changes in color vision, due to a lifetime of exposure to UV B radiation from the sun, produced characteristic changes in color naming, especially in the cool-colors part of the stimulus set. 

Method

We used the HRR pseudoisochromatic plate test (Bailey, Neitz, Tait, & Neitz, 2004), which is a quick, well-validated measure of the participant's ability to discriminate between certain colors and gray, and the D-15 and FM-100 arrangement tests, which are based on the similarities in appearance between colors (Birch, Chilsolm, Kinnear, Pinckers et al., 1979). The tests were administered with AI interpreting, using standard procedures. The participants were the same individuals as in Experiment 1, and their color vision data were collected in the same experimental sessions as the color naming data. All subjects but two completed all three color vision tests. One subject did not complete the HRR plates because that test was not available that day, and one subject did not complete the FM-100 because she could not manipulate the samples. We used advanced analysis methods established by others (Bowman, 1982; Knoblauch, 1987; Vingrys & King-Smith, 1988) to extract quantitative statistics from the qualitative results of the arrangement tests (Appendix C). 

Results

Somali participants had worse FM-100 mean error scores than U.S. observers of comparable ages in Knoblauch et al. (1987); t(40) = 9.61, p < 0.0001 (Figure 10A). The Somali FM-100 data were generally more elongated (Figure 10B): Modulation scores were larger, t(40) = 21.1, p < 0.0001, in the (vertical) tritan direction (Figure 10C) than the U.S. data were, consistent with a Type III defect. Similarly, Somali participants made significant errors on the D-15 test (Figure C1A), and their D-15 total color difference scores (Bowman, 1982) were worse than those of the comparably aged British participants of Bowman, Collins, and Henry (1984), t(41) = 4.57, p < 0.0001, with the error scores concentrated in the blue-yellow (tetartan) direction (Figure C1C). However, participants' errors in the arrangement tests were not generally reflected in their HRR plate results. Twenty-eight participants made no errors on the HRR plates, and with three exceptions the remaining participants' errors on the HRR plates were minor (Appendix C). Two exceptions were male participant 7, who was classified as a moderate protan, and male participant 29, who was classified as a severe deutan, based on concordant results on their HRR, D-15, and FM-100 hue test data. Taken together, the color vision tests did not identify inherited color defects among any of the other participants. The third exception was participant 2p, a 71-year-old male who was scheduled for cataract surgery 3 days after testing. He missed two screening plates for blue-yellow defects as well as three tetartan plates, suggesting that he had a Type III color vision defect. His arrangement tests were also suggestive of a Type III defect (Figure 10, blue disks). Unfortunately, he was lost to follow-up. Because he showed no evidence of an inherited color vision deficiency, his data were included in all the data sets in this study. 

Figure 10

 

Age and FM-100 results at 2400 lx compared with data from Knoblauch et al. (1987) on U.S. observers at 1800 lx. (A) Mean error scores of Somali observers (hatched bars) compared with U.S. observers (gray bars). (B) Diagrams of error scores averaged within decade age bins for the present Somali observers (left) and Knoblauch et al.'s (1987) U.S. observers (right). The axes in the top left diagram of panel B are otherwise suppressed throughout; fiducial circles in the left panels show perfect FM-100 scores. (C) Orientation of the averaged data from panel B compared with ranges for U.S. protan, deutan, and tritan observers (shading) plus the nominal tetartan orientation taken from the D-15 test (dashed line). Somali data show mean error score and orientation values consistent with Type III color vision deficits, but there is little evidence of progression over the adult age range.

As outlined in the Introduction, the second aim of this project was to determine whether there was an effect of color vision on color naming, such as that which might occur if Somali participants acquired over time a Type III color vision deficiency due to the phototoxic effects of UV B–rich tropical sunlight on the eye. Therefore, we looked for an effect of age on color vision. The high error scores on the FM-100 test and their general orientation in the tritan direction (Figure 10A, B) are consistent with the participants' high lifetime exposure to UV B. However, the FM-100 angle, mean error score, and orientation score were not associated with age: on a multivariate analysis of variance (ANOVA), Wilks's lambda = 0.939, F(3, 37) = 0.801, p = 0.502. Particularly, the trend for older Somali participants to have higher FM-100 mean error scores (Figure 10A, hatched bars) was not statistically significant—multivariate ANOVA: F(1, 39) = 1.566, p = 0.654 after Bonferroni correction for three comparisons—and there was no significant trend for the modulation of the data (deviations from circular Figure 10B) or their direction of orientation (Figure 10C) to change with age. Similarly, nonsignificant results for age were obtained for the three scores from the D-15 test: Wilks's lambda = 0.869, F(3, 38) = 1.910, p = 0.144. These results show that, over the age range we examined, there was no significant or obvious tendency for older participants to show a more severe Type III color vision deficiency than younger participants. If there was an effect of age on participants' clinical color vision test results, that effect was already present even in participants in their 20s. 

We also looked directly for any tendency for informants' motifs to be related to their color vision test results. None of the participants except for participant 2p made any blue-type errors on the HRR plates, and, more generally, there was no association between participants' motifs and the number of HRR plate errors they made; one-way ANOVA: F(3, 38) = 0.339, p = 0.797. There was also no association between participants' motifs and the orientations of their arrangement test results; multivariate ANOVA: Wilks's lambda = 0.82, F(6, 72) = 1.234, p = 0.299. This indicates that individuals whose arrangement test scores were closest to the tritan orientations were no more or less likely to use a term that glossed to BLUE than were individuals whose test scores were oriented in other directions. Participant 2p, who had a significant color vision anomaly due to cataract, named colors under the Green-Blue motif, which is the opposite of what is predicted by the brunescence hypothesis. Thus, there is no evidence from Experiments 1 and 2 that the individual variation in motif usage and individual variation in color vision were related. 

Informants and their motifs

The third objective of this research was to examine the question of whether more general language change could be the main explanation for the observed diversity in color naming among our informants. Our data could provide empirical evidence on this point if there were an association between the demographic variables of our informants and their color naming motifs, as suggested by Labov (1972). 

We addressed this issue by relating informants' motifs to their demographic data. First, we examined the demographic data to identify distinct nominal and continuous variables that were not highly related to one another (Appendix D). This analysis revealed distinct nominal variables for gender, occupation, and the geographic location of the informant's home in Somalia and continuous variables for age and the number of years the informant spent in the United States. There was little variability across informants in the amount of time spent outdoors, and we do not analyze those individual data further. 

Gender was related to age, t(40) = 0.014, and to the number of years the informant had spent in the United States, t(38) = 0.002 (Figure 2A, B). These results indicate that the men in this study were generally older and had been in the United States longer, whereas the women were younger and had immigrated more recently. However, age and years in the United States showed only a nonsignificant trend to be related to each other (r = 0.278, p = 0.082). 

Next, we examined the data for possible associations between the informant's motif and his or her other demographic variables. Fisher's exact test revealed a statistically significant association between motif and gender (Figure 11A; p = 0.012), but not locality (p = 0.139) or occupation (p = 0.255). Separately, we examined the motifs for possible effects of age and the number of years in the United States using nominal regression. There was a statistically significant effect of age (χ² = 8.689, p = 0.034) but not of years spent in the United States (χ² = 4.729, p = 0.193; Figure 11B). Particularly, coauthor AI had been in the United States for 12 years when he participated in the study. He was perfectly fluent in U.S. English, but his data set was squarely in the Grue motif (Figure 8). Combining the results of these analyses, female participants were younger and all but one 64-year-old used the Green-Blue motif (Figure 11C), whereas male participants were older and used all four motifs (Figure 11E). Because of the significant association between age and gender, it is not possible to further disentangle their effects on motif usage. However, because there was no effect of the years spent in the United States (Figure 11B, D, and F), there is no indication in the present data set that exposure to the English language has increased informants' color lexicons—for example, by increasing the use of loanwords such as buluug (blue) or barbal (purple). 

Figure 11

 

The effects of gender, age, and years in the United States on motif membership. G-B = Green-Blue motif. (A) The fraction of males and females among the informants using each motif. (B) Informant age at the time of testing, and years residing in the United States by motif. (C) Fraction of women who used each motif, by age. (D) Fraction of men who used each motif, by age. (E) The number of years in the United States for women. (F) The number of years in the United States for men.

General discussion

Somali color naming motifs

This study was designed prospectively (a) to determine whether the motifs that Lindsey and Brown (2009) found in the WCS would replicate in a study of a modern African language and (b) to look for an effect of participants' color vision on their naming of colors. We also collected demographic data to search for correlates of color naming that might give us clues about where the individual variability might come from. 

Color naming diversity and color naming motifs

The color lexicon of 41 Somali-speaking informants included 103 monolexemic color terms, which were glossed into the 11 color categories that were previously identified for the WCS (Kay et al., 2010; Lindsey & Brown, 2006; Regier et al., 2005). In Somali, as in English (Lindsey & Brown, 2014) and Hadzane (Lindsey et al., 2015), the striking diversity in color naming reveals itself in two ways: (a) Many terms are used to name the same or similar colors, and (b) the same names are often used in fundamentally different ways and thus, we infer, have different meanings for different individuals. This result is similar to the diversity reported by Lindsey et al. (2015), where Somali color naming data were collected using many fewer color samples (23 samples as opposed to the 145 samples used here) and where “I don't know” was an allowed color term. 

Despite this diversity, the deployment of Somali color terms is nonetheless lawful and consistent when examined in the context of universal patterns of color naming observed in the WCS (Lindsey & Brown, 2006; Regier et al., 2005). When combined with the WCS data, Somali participants' data sets were identified with the four motifs in the WCS (Lindsey & Brown, 2009). When the Somali data were analyzed separately from the WCS, cluster analysis revealed the WCS Green-Blue and Grue motifs, plus a third Neutral cluster that comprised participants who fell into the Dark and Gray WCS motifs. The regularity in Somali color naming is well captured by the 11 glossed color categories and two to four of the motifs revealed by our previous analyses of the WCS (Lindsey & Brown, 2009). These results show that Somali joins U.S. English as two languages in which multiple motifs were revealed by studies that were prospectively designed to look for them. Thus, within-language diversity in color terms (Bimler, 2007; Davies & Corbett, 1994; Davies et al., 1992, 1997; Heider, 1972; Levinson, 2000; Lindsey & Brown, 2009; Lindsey et al., 2015; Özgen & Davies, 1998; Webster & Kay, 2005, 2007) and in the structure of the motifs in which they are embedded (Lindsey & Brown, 2009) is the rule in world color naming systems. This individual variation is broadly consistent with the view that increasing consensus about what the basic color categories are and what terms should name them is an important component of color term evolution. 

Lack of any color vision deficiency effect on color naming

The second goal of this project was to determine whether there was any evidence that participants' color naming was related to their color vision. The data addressed this issue in two general ways. First, Somalia straddles the equator. All participants in the present study lived between −0.4° and +10° latitude in their youth, and they all reported spending most of their waking hours outdoors. Therefore, they experienced the highest possible dose of UV B during their formative years. If the UV B hypothesis were correct, then Somali-speaking informants should have high rates of blue-type color vision anomalies, and many should name the WCS colors without the use of blue. These predictions generally were supported by the present data set: As a group, the color vision of the Somali participants of all ages showed statistically significant blue-type errors, and many Somali participants used Grue, Gray, or Dark motifs in their color naming. 

Second, if the use of the Grue, Gray, and Dark motifs were due to the difference between blue and green being inconspicuous because of an acquired color vision anomaly, there should be an association between individual color vision test results and individual motif usage. This association should be a consequence of the degradation of color vision over the lifetime of the individual due to ongoing exposure of the eye to UV B radiation. Younger people (presumably with relatively clear ocular lenses and intact S-cones) should show relatively normal results on the color vision tests and should use the Green-Blue motif. In contrast, older people (presumably with brunescent lenses and possibly damaged S-cones) should frequently show Type III color vision deficiency and should use the Grue, Gray, or Dark motifs. Somali informants showed D-15 and FM-100 results that ranged from normal to dramatically poor, so there was plenty of variability in the color vision data set to reveal a correlation between color vision test results and age across individuals, if such a statistical association existed. Contrary to that prediction, some degree of Type III discrimination loss was already in place among the youngest informants, but there was no evidence that older informants had greater discrimination loss than younger informants. This result was quite different from what Hardy et al. (2005) found for younger and older Californians. Although there was a strong effect of age on motif usage among our Somali informants, there was no association across individuals between quantitative color vision test results or qualitative color vision diagnosis and color naming. Therefore, whatever was responsible for the variation across Somali individuals in their naming of colors was not closely related to individual variation in their clinical color vision in the short-wavelength part of the visible spectrum. Particularly, participant 2p had a functionally significant blue-type color vision deficiency that was traceable to his cataracts. Contrary to the prediction of the color vision hypothesis, he used the Green-Blue motif, not the Grue, Gray, or Dark motifs, when he named colors. Thus, the UV B hypothesis is broadly consistent with our group data, but it fails as an explanation of individual variability in color naming. 

Language change

The color terms in the Somali language clearly have undergone significant change since de Larajasse's dictionary was published. The color naming data reported here are consistent with most theories of color term evolution, which posit that the earliest color lexicons contained terms for only a few color categories—generally BLACK, WHITE, and RED (Kay & Maffi, 1999). There is evidence that the modern Somali terms for BLACK (madow) and WHITE (cad, cadaan, and cadays) are quite old. They are native color terms, and they are listed as abstract color terms in both de Larajasse (1897) and Abraham (1964; see Table 1). There has also apparently always been a native term for RED, although the older sources list RED as cas or casaan, whereas in our main data set all informants used guduud. Madow, cad, and cas are identified as basic color terms by Berlin and Kay (1969) and Maffi (1990). These results all suggest that the Somali language had a smaller color lexicon as recently as 117 years ago. 

The historical dictionaries reveal that new Somali color terms often came from words for significant items in the environment that were repurposed as abstract color terms, much as Levinson (2000) suggested. There is also evidence in our main data set that this repurposing is far from complete today. Colors terms other than those for BLACK and RED, including some terms for WHITE, had other noncolor meanings in the past, and those categories currently have multiple terms in common use. The popularity of color terms beyond the first three falls off approximately linearly with the order of their popularity. This approximately linear decline (Zipf's law) is common for ordinary words in large corpora in many languages. Here, it suggests that many of these other color terms are still ordinary words used metaphorically to refer to their corresponding colors (Levinson, 2000). Since Abraham's dictionary was published in 1964, additional color terms have been imported into Somali as loanwords from Italian or English, and they too fall along the linear function in Figure 7. Thus, both the historical dictionaries and the popularity statistics we report here suggest that the Somali color lexicon is undergoing change by adding loanwords and noncolor terms while retaining multiple synonymous color terms for many colors. 

We examined our data set for demographic correlates of color term usage that might be consistent or inconsistent with language change as an explanation of the diversity of motifs in the Somali language. The first possible explanation that comes to mind is that informants might be adopting color terms or color meanings from contact with the surrounding U.S. culture. This is clearly not the case because there is no statistically significant association between the number of years informants had spent in the United States and the likelihood that they used the Green-Blue motif that is so prominent in U.S. English (Figure 11B, D, and F). A second possibility is that there might be distinct dialects of Somali, associated with different geographical locations, and possibly different motifs. This possibility is ruled out because all informants spoke the standard dialect of Somali, and there was no association between informants' hometowns and their motifs. 

A third explanation is that color terms are changing as part of more general evolution of the Somali language. It is well known that language change in U.S. English is associated with informant gender and age (Labov, 1990). When new pronunciations, vocabulary, usage, and grammatical forms come to be adopted by the larger society, women under the age of 40 years are in the vanguard (Labov, 1972; Tagliamonte & D'Arcy, 2009). Recent work has extended this finding to speakers of Arabic, where men tended to use the conservative, classical pronunciations and women tended to speak using the informal, rapidly changing pronunciations (Haeri, 1995). In the present data set all women under age 63 years used the Green-Blue motif, and the only woman who used a non–Green-Blue motif was a 64-year-old who used the Grue motif. Although it was not possible, statistically, to disentangle the effects of gender and age, each of these general trends suggests that color naming is changing in the Somali language in tandem with more general language change, with young women leading the way. 

The children of the Somali people who have immigrated to the United States generally speak English as their native language, so empirical diachronic data on change in the Somali language will never be available in the United States. We hope that, as the political situation in Somalia stabilizes, future color naming data will establish whether the trends we have observed in this snapshot in time are continuing in Somalia itself. We speculate that there will be great change in Somali color naming in future years as some bilingual young people from the Somali diaspora return to their homeland, bringing with them new color terms and new ways of understanding color. The relative contributions of these returning bilingual speakers of the Somali language, and the language change happening in Somalia among those who remained, will provide the basis of an interesting study of the effects of culture on the naming of colors and of language change more broadly. 

Acknowledgments

This work was supported by grant BCS-1152841 from the National Sciences Foundation. We are grateful to the participants for their help. 

Commercial relationships: none. 

Corresponding author: Angela M. Brown. 

Email: [e-mail]. 

Address: College of Optometry, Ohio State University, Columbus, OH, USA. 

Footnotes

¹  Here, a color term in all uppercase refers to a universal glossed color category, and a color term in lowercase refers to the color and its colloquial meaning. Title case refers to the motif. Thus, a person might say red when he identifies a red sample in the RED category, and his motif might be Green-Blue.

Footnotes

²  In modern Somali, the voiced pharyngeal fricative is spelled with the letter C. de Larajasse (1897) used an apostrophe (indicating a glottal stop), and Abraham (1964) did not mark that consonant at all.

Appendix A: Stimulus value tables

Table A1

 

Colors in the pilot stimulus set. Samples are listed in the order of their presentation. Munsell designations are from the Munsell Book of Color, Glossy Edition (Munsell Color Services). WCS designations are the row (letter A . . . J) and column (number 1 . . . 40) within the diagram in Figure 3A, B.

 

Table A2

 

Colors in the main stimulus set. Conventions are as in Table A1.

 

Appendix B: Details of the proximity analyses of the motifs

For our motifs analysis, we analyzed the 83-sample pilot data and the 145-sample final data separately. The 145-sample data were analyzed using spectral clustering analysis, as discussed in the main text. Then, proximity analysis was used to identify the motif membership of each informant's data set from the pilot experiment by finding the WCS motif whose centroid was the closest to the informant's data. 

We classified each pilot informant's 83-sample glossed color naming data according to the centroids of the four motifs, , derived from our previous k-means cluster analysis of the WCS (Haeri, 1995; see Figure 1). Each centroid (m = 1 . . . 4) characterizes, for the group of WCS informants assigned to that motif, the relative frequencies in the deployment of the glossed color terms (falling into one of i = 1 . . . 11 universal categories) across the color samples (n = 1 . . . 330) in the WCS color chart. We used these centroids, adjusted for n = 83, to assign each pilot informant's color naming data to one of the four WCS motifs. Classification was performed using a modified Jacaard similarity metric to determine which of the four motif centroids, , was most similar to each informant's pattern of glossed color names (for further details see Lindsey & Brown, 2009).  

To compare all of our results quantitatively to the WCS, we used this same proximity analysis to determine how close each participant (from our 83-sample pilot data set, our 145-sample main data set, and the WCS) was to his or her motif centroid. These proximity values were then used to determine how close our Somali informants were to their respective centroids, compared with the speakers of the 110 WCS languages, on a percentile basis. 

Appendix C: Color vision test results and their analysis

Arrangement tests

We quantified the results of the D-15 test using the method of Vingrys and King-Smith (Lindsey & Brown, 2009). This yielded three parameters: the color angle of the test results (within a color space defined by those authors, where the angle was related to the type of color vision defect), the C-index for the severity of the color vision defect (the parameter was 1.0 for a normal participant and high if the participant showed many errors consistent with a single color vision defect), and the S-index for the scatter of the results (the parameter was high if the results were nearly random). We also calculated Bowman's (1982) total color difference score in order to compare the results of our Somali participants to the British participants of Bowman et al. (Bowman, 1982). 

We quantified the results of the FM-100 test using the method of Knoblauch (1987), which yielded the angle of the error diagram's orientation (which was related to the type of defect), an average error score (an index of the severity of the color vision deficiency), and the modulation of the data (which was low if the subject's errors were nearly random). 

A correlational analysis of the continuous color vision statistics revealed good reliability across the arrangement tests in the severity of defect: The C-index of the D-15 and the mean error score of the FM-100 were significantly correlated across informants (r = 0.489, p = 0.001). Figure C1A and B shows that informants who tested as normal on the D-15 test (red symbols) had low FM-100 mean error scores compared with those who made errors; t(39) = 3.050, p = 0.004. For the FM-100 hue test, the mean error score correlated with the modulation index (r = 0.402, p = 0.009, data not shown), with informants with low total error scores generally showing lower modulation values. For the D-15, the three measures (angle, C-index, and S-index) were correlated with one another (all r > 0.548, all p < 0.0005). 

Other aspects of the arrangement test results did not correlate so well. The S-index for scatter of the D-15 and the modulation index of the FM-100, which indicate the regularity of the data, were not correlated with one another (r = 0.029, p = 0.736). The orientation of the FM-100 results was not different for those who tested as normal on the D-15 compared with those who made D-15 errors (red vs. white symbols in Figure C1D); Levene's test for equality of variances: F = 0.034, p = 0.854; t(39) = 0.439, p = 0.663. Using the orientation parameters for the FM-100 and the D-15 results, we arrived at a diagnosis for each participant for each color test (Figure C1C, D). We used the orientations for each FM-100 diagnosis from Knoblauch (1987), and we added a tetartan pattern, which we defined using the same angle in CIE₁₉₃₁ color space as that specified for the D-15 test. There was no clear association between the diagnoses based on the D-15 test (protan, deutan, tritan, tetartan, normal, or other) compared with the FM-100 test (protan, deutan, tritan, tetartan, or other; Fisher's exact test, p = 0.794). 

Figure C1

 

Distributions of color vision test results. Red symbols represent results for participants who tested as normal on the D-15 panel. Blue disks represent results for participant 2p, who had a Type III (blue) color vision deficiency. (A, B) Error index histograms and (C, D) orientation histograms for the D-15 and FM-100 hue test results, respectively. Vertical fiducial lines indicate severities and angles for the possible color vision diagnoses. (A, C) N: normal; D′: deuteranomalous; T′: tetartanope; T: tritanope, D: deuteranope, P: protanope, from typical color-deficient D-15 data. (B, D) FM-100 data from Knoblauch et al., 1987. N: average data on 42 normal observers in their 20s, ±1 SD; T, D, P: fiducial lines are from single subjects +/− confiintervals. The intervals around T, D, and P in panels B and D would probably have been larger had multiple individuals been tested. T′ is the tetartanopic angle from the D-15 test, transferred colorimetrically to the FM-100 test.

HRR plates

Twenty-eight participants passed all the plates of the HRR test. Twelve participants failed one or more red-green screening plates but passed all the diagnostic plates. Two participants failed one or two screening plates and one red-green diagnostic plate. Neither of these participants had remarkable findings on the arrangement tests, and their data were retained. One participant (participant 2; see below) missed three screening plates and all four blue-yellow plates; his data were also retained. Overall, there was no statistical association between the diagnostic category from the arrangement tests and the HRR plates missed (blue-yellow, red-green, or none); Fisher's exact test: D-15, p = 0.305; FM-100, p = 0.935. 

We determined that participant 2p had a Type III color vision deficiency. He reported that he had cataracts, and he was scheduled for surgery 3 days after he was tested. He missed screening plates 5, 6, and 7 and all four of the diagnostic plates designed to detect a blue defect. His D-15 and FM-100 results were in the tritan range, his D-15 error C-index was moderate, and his FM-100 mean error score was higher than those of all the color-normal participants. 

Excluded subjects

Two informants had red-green color vision deficiencies (Figure C2) that were identified by their color vision tests results. Their data have been removed from all analyses. 

Figure C2

 

Color naming diagrams and color vision arrangement test results for the two color-deficient participants. Fiducial lines on the D-15 diagram are identified as protan, deutan, tritan (“Tr”), and tetartan (“Tt”). Dashed lines through the FM-100 diagrams are the orientation scores, calculated by the method of Knoblauch (1987). (A) Male participant 7, age 92 years, moderate protan. Notice the blue-type errors on the D-15. (B) Male participant 29, age 87 years, severe deutan. The FM-100 test is shown at greatly expanded scale, with the bold line in the middle range being the error score = 14 ring, the outermost ring in panel A. Notice the red color terms scattered among the green samples.

Appendix D: Statistical analyses of demographic data

Time outdoors

Estimated time outdoors was available for 32 informants. This was a highly skewed distribution, with 29 informants (88%) spending at least 75% of their time outdoors. Participants' time spent outdoors was unrelated to their gender, occupation, locality, age, or years spent in the United States. 

Age and years in the United States

The participant's ages were highly correlated with their ages at immigration but not with the number of years they spent in the United States. Therefore, we eliminated their ages at immigration from our analyses. 

Occupation and locality

Participants' occupation and the locality within Somalia were unrelated to each other by Fisher's exact test (p = 0.072). Age was highly associated with occupation, F(3, 32) = 4.956, p = 0.007, because people who worked in the home (as homemakers or as children and students) were younger than business people (p = 0.027), farmers (p = 0.009), or herders (p = 0.027) on post hoc analyses with Tukey's honestly significant difference correction for repeated comparisons. There was no association between the number of years a person had spent in the United States and his or her occupation in Somalia, F(3, 32) = 1.717, p = 0.185. 

Gender

The continuous variables for age and the number of years the informant had spent in the United States were both related to the informant's gender; age: t(36.3) = 2.693, p = 0.011; years in the United States: t(27.4) = 3.296, p = 0.003. However, gender was not significantly associated with occupation (Fisher's exact test: p = 0.107). Children of both sexes were at home as students, thus overriding the fact that homemakers were all women, and equal numbers of men and women engaged in the other occupations (business, farming, herding of animals).