Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.psy.msu.ru/people/izmailov/1992-izmailov_sokolov.pdf Дата изменения: Fri Oct 21 22:41:50 2011 Дата индексирования: Tue Oct 2 00:16:12 2012 Кодировка:

PSYCHOLOGICAL SCIENCE

Research Article
A SEMANTIC SPACE OF COLOR NAMES Ch.A. Izmailov and E.N. Sokolov
Moscow State University Russiansubjectslearnedarbitrary Abstract Three pairingsbetween20 colors and 20 three-letter artificialcolor names. After amountsof this training,the subjectsrated the differdifferent ence 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 to clusterscorresponding the following groups of related colors: the violets, the blues, the greens, and the yellows-throughreds. After more extensive training,multidimensional scaling yieldedthefull color circle of hues. Furtheranalysisof the data indicatedthat a spherical model previouslyproposed by the authors sensory color space has advantages, also, for the for semanticcolor space obtainedwhen only the names of colors are presented. The results are interpretedin terms of a twostage process ofneuronal analysis of visualinputsin whichthe channels is followed by differactivityof four color-opponent entialactivationof cells tuned to specific colors. of The construction a semanticspace of color namesattracts of scientists'attentionas muchas the construction a subjective spaceof color stimuli.The resultsof applyingvariousmethods, such as the semanticdifferential,factor analysis, and multidimensionalscaling, can be summedup as follows: to 1. The scalingof basic color names,mostlythose relating 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 experimental way 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-C420 color TV sets controlled by CM-1403computer. Colors were displayed on one of the screens, words on the other. The colors over the spectrumbetweenviolet and red, and were distributed 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 experimental procedureconsistedof three stages. First, wordsthat have no color meaning we pickedout 20 three-letter for Russian-speaking subjects. 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 sensory experience informationon the changes in the initial semantic structure name is determinedby the corresponding offeredby an individual'svisual system. The semantics of a broughtaboutby color experience. color name includes, besides visual experience, some material 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, StateUniver- the subjectestimatedthe differencein color betweenthe words' Moscow to Address correspondenceCh.A.Izmailov, of sity, 18/5Department Psychology,103009,Moscow,Prospekt meaningson a scale from0 (completeidentity)to 9 (maximum difference).Each word in a pair was shown for 0.8 s, with an Russia. Marxa, 1992 VOL.3, NO.2, MARCH American © Society Psychological Copyright 1992 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 of everypossibletwo-wordcombination the 20 wordson the list was random. had appeared10 times. The orderof presentation matrixof 20 (20 Foreach subject,we obtaineda triangular 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-dimensional euclideanspace, and then the coordinates the of the 20 pointsrepresenting 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 dimenin sions are summarized Table 1. used in the thirdphasewas identicalto thatin The procedure the first phase. We againobtaineda matrixof differenceestimates for every pair of words and analyzedit by metricmultidimensional scaling.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 correlationcoeffiof cient with number dimensions.This patternindicatesthatthe selected artificialwords were initiallydevoid of color semantics. The results of the multidimensional scalingof semanticdifoffer an entirelydifferentpicture.There ferences aftertraining 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 The projectionsof the 20 points dimensionsare unimportant. artificialcolor names onto the XXX2 plane of the representing euclideanspace aregiven in Figure1. The pointsare distributed aroundfour loci at some distancefrom each other. Each locus the comprisespoints representing namesof colors fromone of the four differentpartsof the visual spectrum. For example, in Figure la, the names of colors from the indigo-blue partof 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 neighboring colors, 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 fournatural color 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 individual features). The obtainedconfiguration the pointsdiffersconsiderably of fromthe color circle. The questionarose whetherwhat we obtained is the final or an intermediate structureof 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

Phase 3

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 corresponding word would appearon the screen of the other TV display with a slight delay. The colorword pairs were presented five times each in randomorder. after this training,the subjects were shown only Immediately the colorsandaskedto nameeach by its corresponding artificial 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 thirdexperimental phase the buildingup of a semanticspace of artificialcolor names.

Table1. Characteristic roots and coefficientsof correlationobtainedin the multidimensional scaling analysis of the matricesfor two subjects in Phase 7, before training Characteristic root L.S. 7667 4118 3195 2783 2610 2449 L.Sh. 13330 9504 8103 5824 4515 3715 Coefficient of correlation L.S. .52 .64 .66 .70 .72 .77 L.Sh. .52 .66 .77 .81 .87 .89

Dimension of space 1 2 3 4 5 6

Note.Results thefirstsix dimensions arepresented. for only

106


PSYCHOLOGICAL SCIENCE

Ch.A. Izmailov and E.N. Sokolov
Table2. Semanticdifferencesbetweenartificialwordsobtainedfor two subjects after Phase 2 training Word 4>HP EYM BAn TAH flAX XAfl KMB JIYC fl3K XAIJ >KOK 3AH MEK HK»K IIEB PYH CAB TOJI EH0> flMJI No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 34 30 48 58 78 90 78 90 74 78 90 86 58 46 26 34 10 10 60 2 66 10 56 46 58 64 66 66 68 60 78 70 56 54 32 40 32 30 54 3 62 24 48 54 64 60 64 58 62 74 84 80 70 50 34 38 36 34 54 4 54 16 10 10 34 30 32 30 36 68 80 74 62 58 62 54 62 60 56 5 66 18 14 12 36 30 44 34 42 70 80 78 66 60 60 54 58 64 54 6 66 16 16 14 10 22 20 24 18 70 86 82 70 66 78 68 76 78 44 7 82 76 68 76 76 74 14 10 28 72 88 84 76 68 76 60 82 86 50 8 86 78 76 74 74 72 14 14 28 72 82 80 70 72 76 64 76 78 48 9 80 74 76 70 70 68 12 10 24 74 86 88 72 72 80 62 62 84 48 10 80 82 84 76 76 68 40 24 30 34 74 78 44 62 74 52 72 74 28 11 86 82 80 78 80 82 72 70 74 72 28 23 34 52 66 62 74 72 34 12 86 86 84 82 80 82 78 80 82 76 28 10 68 72 80 74 78 82 56 13 88 82 80 84 82 80 80 76 76 68 12 10 70 70 74 74 82 82 58 14 74 84 82 78 80 80 82 80 78 80 76 82 76 26 48 30 54 58 30 15 52 68 72 66 68 80 80 76 80 78 78 82 82 28 46 20 38 48 42 16 34 68 76 76 74 80 80 80 80 82 80 78 80 26 24 28 30 48 48 17 50 68 68 70 72 74 80 84 74 84 80 82 80 46 36 32 22 34 46 18 10 68 68 64 60 64 82 80 78 82 78 84 80 62 40 36 34 18 54 19 10 46 60 70 64 74 80 82 80 80 82 78 80 68 66 56 24 10 60 20 84 84 84 62 72 76 58 28 42 44 76 80 78 76 78 82 80 84 82

across10presentations eachwordpair.Theupper of Note.Datawereaveraged the right-hand triangle presents datafor subject L.S., the andthelowerleft-hand L.Sh. triangle presents dataforsubject dimensional space 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-dimensional space and those for spaces with more than two dimensionsdecreased.This result corrobcolor nameshas only oratesthe claimthatthe space of artificial two dimensions. the of The configuration points representing artificialcolor euclideanspace is shown in Figure namesin a two-dimensional 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 of situatedon the circumference the color circle, in accordwith roots and coefficientsof Table3. Characteristic obtainedin the multidimensional correlation scaling analysisof the matricesin Table2 Characteristic root L.S. 16389 12496 9328 4277 2420 1653 L.Sh. 15236 10802 4186 1987 1668 1559 Coefficient of correlation L.S. .66 .86 .94 .96 .98 .98 L.Sh. .66 .94 .97 .96 .96 .97 the hues of the corresponding colors. 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-color of channels,andthenlearna refinedorganization the words in agreementwith the spectralarrangement the color of 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 traditional euclideanmodels of color vision. The Spherical Model of Color Concept Space the The mainidea underlying sphericalmodel is that in the visualsystem, lightis analyzedby meansof four neuronalchannels: two chromaticchannels (red-greenand blue-yellow)and channels(brightand dark).The channels9 two achromatic outputs 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 correspondsmathematically the equato tion of a sphericalsurfacein four-dimensional euclideanspace. A set of pointson the surfaceof such a sphererepresentsthe set
107

Dimension of space 1 2 3 4 5 6

for Note.Results thefirstsix dimensions arepresented. only

VOL. 3, NO. 2, MARCH1992


PSYCHOLOGICAL SCIENCE Semantic Space of Color Names

to Fig. 1. Semanticspace of artificialcolor namesafterone trainingsession. Pointscorresponding color categoriesformfourloci in approximate accordancewith Hering'sopponentcolors, (a) SubjectL.S. (b) SubjectL.Sh. of colorsdiscriminated the visual system. Each color pointis The threecharactenstics hue, brightness,and saturation by fixed by setting a single combinationof four coordinates,al- are representedin the sphericalmodel as three sphericalcoorcan thoughthe same combination correspondto lightswith dif- dinates.The modelgives a new description,on the one hand,to ferent spectral distributions(Izmailov, 1980, 1982; Izmailov, the interconnection the neurophysiological of channels, which are representedby the four Cartesiancoordinatesof a color Sokolov, & Chernorizov,1989). The two chromatic channelsanalyzethe spectraldistribution point on the sphere, and, on the other hand, to the sensory of light that is perceived as hue. The achromaticchannels an- characteristics light, representedby the three sphericalcoof alyze lightintensity,perceivedas color brightness.Because the ordinatesof the same point. chromatic achromatic and channelsare relatedvia the spherical For equibright lightsfromdifferentpartsof the spectrum,as of law, the combination their activitiesyields anothercolor pa- used in ourexperiments,the subsetof colorsfalls on a spherical rameter saturation,which compensatesto a certainextent for surfacein a three-dimensional euclideanspace. If the structure the loss of physical informationin visual processing. Adding of artificial color namesreflectsthe activityof the sensorysyssaturationdefines third polar coordinates of the four-dimen- tem, then the configuration artificial of color namesobtainedby sional color space. multidimensional shouldbest be represented the suron scaling face of a spherein a three-dimensional space, also, and not on a plane, as in Figures 1 and 2. If so, Figures 1 and 2 give only Table4. Characteristic roots and coefficientsof the projectionsof the points onto the XXX2 plane. A thirddicorrelationobtainedin the multidimensional scaling mension would be necessary to represent the configuration analysis of the matricesfor two subjects after a second fully. trainingsession we By testing the sphericityof the configurations, 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. euclidean space, we found the L.Sh. L.S. L.Sh. scaling in a three-dimensional centerof the sphere the theoreticalpointas nearlyas possible 1 15766 15010 .72 .74 fromall the points. Since the datacontainerror,the equidistant 2 12715 13094 .89 .94 sphericallayer has a certainthickness,definedas the standard 3 8360 3800 .95 .96 deviationof the radii. On each iteration,a program calculates 4 5086 2230 .97 .97 the radiiof a spherewith a given center, the meanof the radii, 5 1693 1484 .98 .96 and the standard deviation.The program then adjuststhe posi6 1354 1248 .98 .97 tion of the centeruntilthe standard deviationis minimum. For Note.Results thefirstsix dimensions arepresented. for a three-dimensional only space, the solution is consideredacceptable if the thicknessof the spherecontainingthe experimental 108 VOL.3, NO.2, MARCH 1992


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 resemblesNewton's color circle, which is typicalfor colors in sensory space, (a) SubjectL.S. (b) SubjectL.Sh. configuration pointsdoes not exceed 10to 12%of the meanradius(Izmailov, 1980;Izmailovet al., 1989). for The resultsof testingall the experimental configurations are given in Table 5. The thicknesses of the layers sphericity are the points before training 27%and containing experimental 21%,respectively,for subjectsL.S. and L.Sh. Thereis no sigevidenceof the absenceof nificantsphericity.This is additional 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 resultingmultidimensional scaling solution for artificialcolor names, presentedin Figure 3, has the same circular orderof 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 semanticspaces in terms of radii of points Table5. Indices of sphericityof three-dimensional artificialcolor names representing Before learning Index M SD CV
r

After 1 run L.S. 44.9 2.7 6.0
.94

After 2 runs L.S. 43.2 2.3 6.5
.95

After 5 runs V.L. 37.2 4.2 11.0
.96

L.S. 29.6 8.0 27.0
.66

L.Sh. 38.6 8.1 21.0
.77

L.Sh. 43.4 7.0 16.0
.97

L.Sh. 44.8 6.1 13.5
.96

Note. M = mean radius;SD - standard deviation;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
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)

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

VOL. 3, NO. 2, MARCH 1992