PSYCHOLOGICAL
SCIENCE
Research Article
A SEMANTIC SPACE OF COLOR NAMES
Ch.A. Izmailov and E.N. Sokolov
Moscow State University
- ThreeRussiansubjectslearnedarbitrarypairingsbeAbstract
tween20 colors and 20 three-letterartificialcolor names. After
differentamountsof this training,the subjectsrated the difference betweenthe colors associated with everypair of artificial
names when these names were presented without the colors.
Multidimensional
scaling of the ratingsafter a small amountof
trainingrevealed a groupingof the words into four semantic
clusterscorrespondingto the following groups of related colors: the violets, the blues, the greens, and the yellows-throughreds. After more extensive training,multidimensionalscaling
yieldedthefull color circle of hues. Furtheranalysisof the data
indicatedthat a spherical model previouslyproposed by the
authorsfor sensory color space has advantages, also, for the
semanticcolor space obtainedwhen only the names of colors
are presented. The results are interpretedin terms of a twostage process ofneuronal analysis of visualinputsin whichthe
activityof four color-opponentchannels is followed by differentialactivationof cells tuned to specific colors.
The constructionof a semanticspace of color namesattracts
scientists'attentionas muchas the constructionof a subjective
spaceof color stimuli.The resultsof applyingvariousmethods,
such as the semanticdifferential,factor analysis, and multidimensionalscaling, can be summedup as follows:
1. The scalingof basic color names,mostlythose relatingto the
basic colors of the spectrum(e.g., yellow, green, and red),
yields a two-dimensionalspace of color names similarto
Newton's color circle (Fillenbaum & Rapoport, 1971;
Shmelev, 1983).
2. The inclusionof color names that refer to materialcharacteristics of colored objects (e.g., lemon, emerald, marshgreen, khaki,coffee, and gold) increasesthe dimensionality
and disorganizesthe circular structureof the basic color
space (Artemieva,1968;Sokolov & Vartanov,1987).
An experimentalway of testing this hypothesismightbe to
create a set of artificialcolor names whose semanticsis determined beforehand,and then compare the semantic space of
these artificialcolor names with that of naturalones. This article presents the results of a series of experimentsaimed at
buildingsuch a semanticspace of artificialcolor names.
METHOD
Apparatus and Stimuli
The stimuliwere generatedby two Elektronika-C420color
TV sets controlled by CM-1403computer. Colors were displayed on one of the screens, words on the other. The colors
were distributedover the spectrumbetweenviolet and red, and
also includedwhite and a numberof purplemixturesof short
andlong wavelengths.The subjects'reactionswere put into the
computerthroughan 11-keyminiterminal.
Subjects
The experimentswere performedwith three subjects, aged
22 to 25, with normalcolor vision; two subjects served in the
mainexperimentand one only in an auxiliaryexperiment.
Procedure
The experimentalprocedureconsistedof three stages. First,
we pickedout 20 three-letterwordsthat have no color meaning
for Russian-speakingsubjects. The initial spatial structureof
the chosen wordswas tested by metricmultidimensional
scaling
(Izmailov, 1980;Torgerson, 1958). Our aim was to obtain the
initial backgroundstructureof the meaninglesswords on the
list. Second, each subjectwas then taught,by simple associaThe mainreasonfor the discrepancybetween the results of tive training,pairingsbetween the words and 20 colors (Arscalingthese two types of color names- the basic and the ma- temieva, 1968;Miller, 1969, 1971).Third,the same methodof
terialnames- lies hypotheticallyin the differenceof their se- multidimensionalscaling was used to construct a semantic
manticcontent and structure.The semantics of a basic color space for the artificialcolor names. In this way, we obtained
name is determinedby the correspondingsensory experience informationon the changes in the initial semantic structure
offeredby an individual'svisual system. The semantics of a broughtaboutby color experience.
materialcolor name includes, besides visual experience, some
other types of cognitive experience, especially that of speech
Phase 1
(Artemieva,1968;Miller, 1971).
In Phase 1, pairs of the artificialwords from the main list
were successively presentedon the TV screen. For each pair,
to Ch.A.Izmailov,MoscowStateUniver- the subjectestimatedthe differencein color betweenthe words'
Addresscorrespondence
of Psychology,103009,Moscow,Prospekt meaningson a scale from0 (completeidentity)to 9 (maximum
sity, 18/5Department
difference).Each word in a pair was shown for 0.8 s, with an
Russia.
Marxa,
VOL.3, NO.2, MARCH1992
© 1992American
Society
Psychological
Copyright
105
SCIENCE
PSYCHOLOGICAL
SemanticSpace of Color Names
intervalbetween words of 0.4 s. Two seconds after the first
pair,a second pairwas presented,then a third,and so on, until
everypossibletwo-wordcombinationof the 20 wordson the list
had appeared10 times. The orderof presentationwas random.
Foreach subject,we obtaineda triangularmatrixof 20 (20 l)/2 elements, each element being the arithmeticmean of 10
estimates. Every matrixwas processed by metric multidimensionalscaling.We calculatedthe eigenvaluesof the coordinates
in a 20-dimensionaleuclideanspace, and then the coordinates
of the 20 pointsrepresentingthe 20 words. The axes were numbered in order of their eigenvalues. Subsequently,we calculated the linear correlation coefficients between the initial
differenceestimates and the interpointdistances in the 20 dimensionsof subjectivespace. The resultsfor the first 6 dimensions are summarizedin Table 1.
Phase 2
The mainpurposeof this phase was to trainthe subjectsto
name each of the 20 colors by a single one of the 20 artificial
words on the list. The colors were mixturesof three basic TV
colors with the following dominantwavelengths:red 620 nm,
green535 nm, and blue 485 nm. The colors were mixedin pairs
in such a way that the stimuli included all the colors of the
spectrumplus purples. In additionto chromaticcolors, white
was obtainedby mixing all three TV colors. All colors were
equatedin brightnessto the white by the methodof heterochromaticphotometryat the level of 15 cd/m.
One of the 20 stimuliwould appearon the screen of one of
the TV displays.The correspondingword would appearon the
screen of the other TV display with a slight delay. The colorword pairs were presented five times each in randomorder.
Immediatelyafter this training,the subjects were shown only
the colorsandaskedto nameeach by its correspondingartificial
word. If all the names were given correctly, the trainingwas
terminated.Otherwise,the trainingcontinueduntil the subject
learnedto give the assignednames to all the colors. Only then
didwe pass on to the thirdexperimentalphase- the buildingup
of a semanticspace of artificialcolor names.
Table1. Characteristicroots and coefficientsof
correlationobtainedin the multidimensionalscaling
analysis of the matricesfor two subjects in Phase 7,
before training
Characteristic
root
Coefficient
of correlation
Dimension
of space
L.S.
L.Sh.
L.S.
L.Sh.
1
2
3
4
5
6
7667
4118
3195
2783
2610
2449
13330
9504
8103
5824
4515
3715
.52
.64
.66
.70
.72
.77
.52
.66
.77
.81
.87
.89
Note.Resultsforthefirstsix dimensions
onlyarepresented.
106
Phase 3
The procedureused in the thirdphasewas identicalto thatin
the first phase. We againobtaineda matrixof differenceestimates for every pair of words and analyzedit by metricmultidimensionalscaling.The matrixof semanticdifferencesis given
in Table 2, and the resultingeigenvaluesand correlationcoefficients in Table 3.
RESULTS AND DISCUSSION1
The semantic space of artificialwords before trainingappears randomin its projectionon the first two principalaxes.
Moreover,the decrease of the eigenvaluesof the axes is relatively uniform,and so is the increaseof the correlationcoefficient with numberof dimensions.This patternindicatesthatthe
selected artificialwords were initiallydevoid of color semantics.
The results of the multidimensionalscalingof semanticdifferences aftertrainingoffer an entirelydifferentpicture.There
is a noticeable shift in the eigenvaluesbetween the first two
axes and the rest. This shift implies that the first two axes
determinethe location of the points in space, while the other
dimensionsare unimportant.The projectionsof the 20 points
representingartificialcolor names onto the XXX2plane of the
euclideanspace aregiven in Figure1. The pointsare distributed
aroundfour loci at some distancefrom each other. Each locus
comprisespoints representingthe namesof colors fromone of
the four differentpartsof the visual spectrum.
For example, in Figure la, the names of colors from the
indigo-bluepartof the spectrum(Points7-9) are groupedat the
bottom right, the names of green colors (Points 11-13) are on
the bottomleft, the namesof yellow and orangecolors (Points
14-19)are at the top left, and the namesof violet colors (Points
2-5) are on the top right.
The points inside the loci are not ordered.The subjectconfused the names of neighboringcolors, but not the names of
colors from different loci. The fact that the artificial color
names fall into four distinctclasses based on wavelengthpermitsus to relatethe fournaturalcolor names- red, blue, green,
andyellow- to Hering'sfour basic colors. We can see that the
semantics of artificialcolor names is determinedmainly by
color stimulation.This conclusion holds for both subjects(althoughthe resultsfor each have some individualfeatures).
The obtainedconfigurationof the pointsdiffersconsiderably
fromthe color circle. The questionarose whetherwhat we obtained is the final or an intermediatestructureof the artificial
color names. Therefore, we repeated the Phase 2 traininga
week laterwith the same subjects.This time the subjectswere
muchquickerto learnall the words. Then, once again,we built
a semanticspace of color namesfor each subject(Table4).
It is clear from the data that additionaltrainingdid not
change the dimensionalityof the semantic space of artificial
color names. It only increasedthe metricprecisionof the two1. Tables of the followingdata may be obtainedfrom the authors:
matrixof semanticdifferencesbetweenstimuluswordsbeforelearning
associationswith colors and matrixof semanticdifferencesbetween
stimuluswordsaftera second session of training.
VOL. 3, NO. 2, MARCH1992
PSYCHOLOGICAL
SCIENCE
Ch.A. Izmailov and E.N. Sokolov
Table2. Semanticdifferencesbetweenartificialwordsobtainedfor two subjects after Phase 2 training
Word
No.
1
4>HP
EYM
BAn
TAH
flAX
XAfl
KMB
JIYC
fl3K
XAIJ
>KOK
3AH
MEK
HK»K
IIEB
PYH
CAB
TOJI
EH0>
flMJI
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
34
30
48
58
78
90
78
90
74
78
90
86
58
46
26
34
10
10
60
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
66
62
24
54
16
10
66
18
14
12
66
16
16
14
10
82
76
68
76
76
74
86
78
76
74
74
72
14
80
74
76
70
70
68
12
10
80
82
84
76
76
68
40
24
30
86
82
80
78
80
82
72
70
74
72
86
86
84
82
80
82
78
80
82
76
28
88
82
80
84
82
80
80
76
76
68
12
10
74
84
82
78
80
80
82
80
78
80
76
82
76
52
68
72
66
68
80
80
76
80
78
78
82
82
28
34
68
76
76
74
80
80
80
80
82
80
78
80
26
24
50
68
68
70
72
74
80
84
74
84
80
82
80
46
36
32
10
68
68
64
60
64
82
80
78
82
78
84
80
62
40
36
34
10
46
60
70
64
74
80
82
80
80
82
78
80
68
66
56
24
10
84
84
84
62
72
76
58
28
42
44
76
80
78
76
78
82
80
84
82
10
56
46
58
64
66
66
68
60
78
70
56
54
32
40
32
30
54
48
54
64
60
64
58
62
74
84
80
70
50
34
38
36
34
54
10
34
30
32
30
36
68
80
74
62
58
62
54
62
60
56
36
30
44
34
42
70
80
78
66
60
60
54
58
64
54
22
20
24
18
70
86
82
70
66
78
68
76
78
44
14
10
28
72
88
84
76
68
76
60
82
86
50
14
28
72
82
80
70
72
76
64
76
78
48
24
74
86
88
72
72
80
62
62
84
48
34
74
78
44
62
74
52
72
74
28
28
23
34
52
66
62
74
72
34
10
68
72
80
74
78
82
56
70
70
74
74
82
82
58
26
48
30
54
58
30
46
20
38
48
42
28
30
48
48
22
34
46
18
54
60
across10presentations
of eachwordpair.Theupperright-hand
Note.Datawereaveraged
trianglepresentsthedatafor subjectL.S.,
andthelowerleft-hand
trianglepresentsthedataforsubjectL.Sh.
dimensionalspace of artificialcolor names. The differencebetween the eigenvaluesof the first two axes and of all the other
axes increased, also. The difference between the correlation
coefficientsfor a two-dimensionalspace and those for spaces
with more than two dimensionsdecreased.This result corroboratesthe claimthatthe space of artificialcolor nameshas only
two dimensions.
The configurationof points representingthe artificialcolor
namesin a two-dimensionaleuclideanspace is shown in Figure
2a (for subject L.S.) and 2b (for subject L.Sh.). Comparison
with Figure 1 gives an idea of the specific changes in the semanticspace caused by additionaltraining.The points previously inside each of the four groups of color names are now
situatedon the circumferenceof the color circle, in accordwith
Table3. Characteristicroots and coefficientsof
correlationobtainedin the multidimensionalscaling
analysisof the matricesin Table2
Characteristic
root
Coefficient
of correlation
Dimension
of space
L.S.
L.Sh.
L.S.
L.Sh.
1
2
3
4
5
6
16389
12496
9328
4277
2420
1653
15236
10802
4186
1987
1668
1559
.66
.86
.94
.96
.98
.98
.66
.94
.97
.96
.96
.97
Note.Resultsforthefirstsix dimensions
onlyarepresented.
VOL. 3, NO. 2, MARCH1992
the hues of the correspondingcolors. This patternis especially
clear in the data in Figure2a.
These results suggest that subjectsfirst learn to groupnew
color words into general categories based, perhaps, on opponent-colorchannels,andthenlearna refinedorganizationof the
words in agreementwith the spectralarrangementof the color
circle. Because the formationof the basic color categories is
thus determinedby the activity of the visual system, the semanticspace of basic color names such as blue, green, yellow,
and white correspondsclosely to the sensory color space.
This conclusion can be interpretedmore accurately if we
analyze the results in terms of the spherical model of color
vision (Izmailov, 1980, 1982; Izmailov & Sokolov, 1991;
Sokolov & Vartanov, 1987). The reason is that the spherical
modelimposes morelimitationson the structureof the sensory
color space than traditionaleuclideanmodels of color vision.
The Spherical Model of Color Concept Space
The mainidea underlyingthe sphericalmodel is that in the
visualsystem, lightis analyzedby meansof four neuronalchannels: two chromaticchannels (red-greenand blue-yellow)and
two achromaticchannels(brightand dark).The channels9outputs become the inputs of the color-detectorcells, each tuned
selectively to a certain color determinedby a specific combination of the coefficients of synaptic transmission.Although
each color detector has its own combinationof synaptic coefficients, the sum of the squaresof the coefficients is constant
across detectors.This correspondsmathematicallyto the equation of a sphericalsurfacein four-dimensionaleuclideanspace.
A set of pointson the surfaceof such a sphererepresentsthe set
107
PSYCHOLOGICAL
SCIENCE
Semantic Space of Color Names
Fig. 1. Semanticspace of artificialcolor namesafterone trainingsession. Pointscorrespondingto color categoriesformfourloci
in approximateaccordancewith Hering'sopponentcolors, (a) SubjectL.S. (b) SubjectL.Sh.
of colorsdiscriminatedby the visual system. Each color pointis
The threecharactenstics- hue, brightness,and saturation
fixed by setting a single combinationof four coordinates,al- are representedin the sphericalmodel as three sphericalcoorthoughthe same combinationcan correspondto lightswith dif- dinates.The modelgives a new description,on the one hand,to
ferent spectral distributions(Izmailov, 1980, 1982; Izmailov, the interconnectionof the neurophysiologicalchannels, which
are representedby the four Cartesiancoordinatesof a color
Sokolov, & Chernorizov,1989).
The two chromaticchannelsanalyzethe spectraldistribution point on the sphere, and, on the other hand, to the sensory
of light that is perceived as hue. The achromaticchannels an- characteristicsof light, representedby the three sphericalcoalyze lightintensity,perceivedas color brightness.Because the ordinatesof the same point.
chromaticandachromaticchannelsare relatedvia the spherical
For equibrightlightsfromdifferentpartsof the spectrum,as
law, the combinationof their activitiesyields anothercolor pa- used in ourexperiments,the subsetof colorsfalls on a spherical
rameter- saturation,which compensatesto a certainextent for surfacein a three-dimensionaleuclideanspace. If the structure
the loss of physical informationin visual processing. Adding of artificialcolor namesreflectsthe activityof the sensorysyssaturationdefines third polar coordinates of the four-dimen- tem, then the configurationof artificialcolor namesobtainedby
sional color space.
multidimensional
scalingshouldbest be representedon the surface of a spherein a three-dimensionalspace, also, and not on
a plane, as in Figures 1 and 2. If so, Figures 1 and 2 give only
Table4. Characteristicroots and coefficientsof
the projectionsof the points onto the XXX2plane. A thirddicorrelationobtainedin the multidimensionalscaling
mension would be necessary to represent the configuration
analysis of the matricesfor two subjects after a second
fully.
trainingsession
By testing the sphericityof the configurations,we can answer the followingquestion:On which level of color informaCharacteristic
Coefficient
tion processingin the sensory system does the final shapingof
root
of correlation
basic color categoriestake place? For every multidimensional
Dimension
of space
L.S.
L.Sh.
L.S.
L.Sh.
scaling in a three-dimensionaleuclidean space, we found the
centerof the sphere- the theoreticalpointas nearlyas possible
1
15766
15010
.72
.74
equidistantfromall the points. Since the datacontainerror,the
2
12715
13094
.89
.94
sphericallayer has a certainthickness,definedas the standard
3
8360
3800
.95
.96
deviation
of the radii. On each iteration,a programcalculates
4
5086
2230
.97
.97
the
radii
of
a spherewith a given center, the meanof the radii,
5
1693
1484
.98
.96
and
the
standard
deviation.The programthen adjuststhe posi6
1354
1248
.98
.97
tion of the centeruntilthe standarddeviationis minimum.For
Note.Resultsforthefirstsix dimensions
a three-dimensionalspace, the solution is consideredacceptonlyarepresented.
able if the thicknessof the spherecontainingthe experimental
108
VOL.3, NO.2, MARCH1992
PSYCHOLOGICAL
SCIENCE
Ch.A. Izmailov and E.N. Sokolov
Fig. 2. Semanticspace of artificialcolor names after a second session of trainingthat fixed the results of the first run. The
configurationresemblesNewton's color circle, which is typicalfor colors in sensory space, (a) SubjectL.S. (b) SubjectL.Sh.
pointsdoes not exceed 10to 12%of the meanradius(Izmailov,
1980;Izmailovet al., 1989).
The resultsof testingall the experimentalconfigurationsfor
sphericityare given in Table 5. The thicknesses of the layers
containingthe experimentalpoints before trainingare 27%and
21%,respectively,for subjectsL.S. and L.Sh. Thereis no significantsphericity.This is additionalevidenceof the absenceof
initialstructurein the artificialwords. Trainingbringsabout a
decreasein the variationof the radii,reflectingthe shapingof a
sphericalstructure.
Hering'sopponentcolors. In the second stage, colors are differentiatedin hue accordingto Newton's color circle. Additionalevidence for this conclusioncomes from the data of subject V.L., obtainedafter more extensive (five runs)learningof
color names. The resultingmultidimensionalscaling solution
for artificialcolor names, presentedin Figure 3, has the same
circularorderof pointsas in Newton's circle, andthe sphericity
of this space is most like that found for sensory color space
(Table 5; Izmailov, 1980; Izmailov et al., 1989; Sokolov &
Vartanov,1987).
An analysisof sphericityalso supportsthe idea thatthereare
two stages in the learningprocess. Color-opponentaxes are
Additional Data and Conclusions
formedat the first stage of training.The sphericalstructureof
Color category formationin the course of trainingdivides the semanticspace of colors appearslater, with repetitionand
into two stages. The first stage results in the separationof ar- fixing of word-color associations.
The activityof the sensory system can affect the brainareas
tificial words into four classes perhaps corresponding to
Table5. Indices of sphericityof three-dimensionalsemanticspaces in terms of radii of points
representingartificialcolor names
Before learning
After 1 run
After 2 runs
After 5 runs
Index
L.S.
L.Sh.
L.S.
L.Sh.
L.S.
L.Sh.
V.L.
M
SD
CV
29.6
8.0
27.0
38.6
8.1
21.0
44.9
2.7
6.0
43.4
7.0
16.0
43.2
2.3
6.5
44.8
6.1
13.5
37.2
4.2
11.0
.96
.96
r
.66
.77
.94
.97
.95
Note. M = mean radius;SD - standarddeviation;CV = coefficientof variance(%);r = coefficientof
correlation.
VOL. 3, NO. 2, MARCH1992
109
PSYCHOLOGICAL SCIENCE
Semantic Space of Color Names
color names in the semantic space, that is, the structure of color
categories.
In conclusion, the semantic space of artificial color names
representing color stimuli without other semantic links closely
corresponds to sensory space. The similarity of the semantic
space of artificial color names and the semantic space of natural
color names supports the hypothesis that the dominant constituent in the semantics of natural color names is the sensory one.
Our hypothesis is that the formation of color categories is a
result of the activity of color-opponent channels and colordetector cells. The dynamics and the direction of the influence
of these two factors on the shaping of the semantic color space
differ substantially.
REFERENCES
Fig. 3. Semantic space of artificial color names after repeated
training and fixing. Subject V.L.
engaged in the semantic coding of information along two independent paths. The first path probably begins in the striate cortex, in the ending zone of the axons of relay cells of the lateral
geniculate body. Here are formed the output signals of the
color-opponent channels of the color analyzer that are responsible for the quick formation of the basic axes of the semantic
space. The second path begins, presumably, in the poststriate
areas of the sensory system, in the v2-v4 zones. From here
originate the output signals of the color-detector cells responsible for the gradual shaping of the spherical structure of the
110
Artcmieva, Ye.Yu. (1968). Subjective semantic psychology. Unpublished doctoral thesis, Moscow State University, Moscow, (in Russian)
Fillenbaum, S., & Rapoport, A. (1971). Structure in the subjective lexicon. New
York: Academic Press.
Izmailov, Ch.A. (1980). Spherical model of color discrimination. Moscow: Moscow State University, (in Russian)
Izmailov, Ch.A. (1982). Uniform color space and multidimensional scaling
(MDS). In H.-G. Geissler, H.F.J.M. Buffart, E.L.J. Leeuwenberg, & V.
Sarris (Eds.), Psychophysical judgment and the process of perception (pp.
52-62). Berlin: VEB Deutscher Verlag.
Izmailov, Ch.A., & Sokolov, E.N. (1991). Spherical model of color and brightness
discrimination. Psychological Science, 2, 249^259.
Izmailov, Ch.A., Sokolov, E.N., & Chernorizov, A.M. (1989). Psychophysiology
of color vision. Moscow: Moscow State University Publishers, (in Russian)
Miller, G.A. (1969). A psychophysical method to investigate verbal concepts.
Journal of Mathematical Psychology, 6, 169-191.
Miller, G.A. (1971). Empirical methods in the study of semantics. In D.D. Steinberg & J.A. Jakobovits (Eds.), Semantics: An interdisciplinary reader in
philosophy, linguistics and psychology (pp. 569-585). Cambridge, England:
Cambridge University Press.
Shmelev, A.G. (1983). An introduction to experimental psychological semantics.
Moscow: Moscow State University, (in Russian)
Sokolov, E.N., & Vartanov, A.V. (1987). On the semantic color space. Psikhologicheskii Zhurnal, 7, 58-65. (in Russian)
Torgerson, W.S. (1958). Theory and methods of scaling. New York: Wiley.
(RECEIVED 5/ZW5J1;ACCEPTED 10/21/91)
VOL. 3, NO. 2, MARCH 1992