An Experimental Comparison of Three Natural Language Colour Naming Models

The problems inherent in providing natural language generation of colour names are discussed. Three models for generating natural language descriptions of HSL colours are described. The effectiveness of these models in describing colours is compared experimentally. It is concluded that a rigid syntactic mapping of HSL components to orthogonal linguistic axes is generally inferior to exhaustive enumeration of colours or custom selection of adjectival colour modifiers. Interesting variations of model preference for different hues and different numbers of adjectival modifiers are noted. Introduction Research in colour modelling for computational purposes has concentrated on finding numerical representations of colours which are convenient for use in computer graphics and image processing. These representation schemes use a three dimensional colour space (RGB, CMY, YIQ, HSV, HSL, XYZ, L*a*b*, L*u*v*, etc.) and quantify colours as points in that space [3,4,6,9]. This provides a means of accurately specifying most colours in terms of a small set of parameters which are suitable for subsequent numeric manipulation. Schwarz, Cowan and Beatty [7] conducted an extensive comparison of human colour matching performance using five standard numerical colour models. Their study indicated that considerable experience is required to effectively understand and use numerical colour specifications. The problem is that none of these colour spaces map well onto the ways humans think about colours. Given a random set of RGB or HSL values, even experienced humans can have difficulty determining what colour is being represented. What colour is represented by the RGB triple [ 1.0 , 0.5 , 0.25 ]? What about the normalised HSL triple [ 0.06 , 0.75 , 1.0 ]? The reverse process is perhaps even more difficult. What is the RGB triple which represents the colour "khaki"? What is the equivalent HSL triple ? At Monash University we are developing a multi-media interface which coordinates automatically-generated natural language explanations with synthetic animations of information to be conveyed. It is in this context that the two related problems of generation and comprehension of natural language colour descriptions have arisen. For example, it is easier to understand software which refers to a colour as "light tan", rather than "RGB [ 1.0 , 0.5 , 0.25 ]". Similarly, people generally prefer to say "khaki" , instead of "HSL [ 0.2, 0.4 , 0.95 ]" In 1982 Berk, Brownston and Kaufman [1,2] proposed a semantic classification system called CNS (for "Colour Naming System"), based on the widely used ISCC-NBS system [11]. CNS is a simple grammar which linguistically encodes a quantised HSL space of 627 distinct colours. The study suggested that proficient CNS users could more accurately identify colours than equally experienced HSL or RGB users. This paper proposes two new semantic colour models and compares them with a variant of the CNS model. The intention is to gather evidence regarding the manner in which humans conceive and describe colours they experience and hence draw conclusions regarding the best way to present colours linguistically. The basis of the proposed models is the semantic classification of colours originally represented using a normalised HSL colour space [9]. The normalised HSL model represents a colour by three real values in the range 0 to 1. The first value encodes the colour's dominant hue (H), progressing cyclically from red at 0.0, through green at 0.33, blue at 0.67 and back to red at 1.0. Intermediate values encode composite colours such as yellow, turquoise and purple. The second value encodes the purity or tone of the colour (its saturation S), ranging from no colour (a grey tone) at 0.0, to a vivid tone of the colour (no grey) at 1.0. The final value encodes the brightness of the resulting shade (its luminance L), from black at 0.0 to full brightness at 1.0. Proposed Colour Models 1. Comparison model The Comparison model is a direct semantic quantisation of HSL space. Each colour name corresponds to a particular point in HSL colour space and to its immediate neighbourhood (that is: the unnamed colours closer to that named HSL point than to any other named point.) In this model "closeness" is defined as Euclidean distance in normalised HSL space, though any other suitable metric could be used. The model postulates that humans use a discrete, empirical, non-uniform colour name space, constructed by comparison with their real-world experience of colour. Colour names in this space are semantic labels for the fuzzy sets of real-world phenomena with which colours are compared. A very small Comparison colour space HSL triple Colour name [ 0.11 , 0.27 , 0.96 ] wheat [ 0.40 , 0.67 , 0.55 ] sea green [ 0.58 , 0.22 , 0.56 ] slate [ 0.62 , 0.64 , 0.67 ] cobalt [ 0.99 , 0.83 , 0.89 ] alizarin crimson Each person's Comparison colour space (CCS) will be unique, reflecting their vocabulary and individual experience of colours in the real world. Colour names in the CCS will be distributed irregularly through HSL space according to the individual. Some people may have a very extensive CCS and will know and be able to linguistically distinguish hundreds of colours. Others may have a comparatively small CCS, knowing and using only a few dozen individual colour names. Significant variations in CCS are also to be expected between cultures. It is also possible that many people will have a larger CCS for recognition than for recall. That is, on seeing a colour, a person asked to select the name of the colour from a list may be able to classify the colour with more discrimination than if they were to attempt the task without prompting. A Comparison colour space can be implemented as a simple list of distinct colour names and associated HSL values. Classification of an arbitrary HSLspecified colour is then simply a process of determining the closest HSL value in the CCS and assigning the associated name to the colour. Example: Find the CCS name for the colour represented by HSL [ 0.3 , 0.3, 0.3 ] Colour name Distance to [0.3,0.3,0.3] wheat 0.68 sea green 0.46 slate 0.39 cobalt 0.59 alizarin crimson 1.05 Hence classify HSL [ 0.3 , 0.3, 0.3 ] as "slate" Obviously, the accuracy with which a colour can be classified will be determined by the number and chromatic distribution of colours in the CCS. For the purposes of this experiment a list of 179 colour names was compiled, with a mean Euclidean separation between nearest neighbours of 0.12 (maximum separation 0.73) 2. Qualification model The Qualification model is a variant implementation of the CNS model. The model hypothesises that most individuals have only a very small Comparison colour space. Alternatively, it suggests that although many people may possess a substantial colour vocabulary, they use only a small subset of it in everyday life. This smaller set of base colour names, perhaps only 10 to 20 in total, represents a crude quantisation of hue (H value). Classification is refined, where necessary, by the use of appropriate adjectives to discriminate between various tones (S values) and shades (L values) of these hues. Qualification colour model (examples) HSL triple Colour name [ 0.11 , 0.27 , 0.96 ] pale bright yellow [ 0.40 , 0.67 , 0.55 ] aqua [ 0.58 , 0.22 , 0.56 ] pale blue [ 0.62 , 0.64 , 0.67 ] blue [ 0.99 , 0.83 , 0.89 ] bright red The model assumes that the list of base colour names is likely to vary only slightly within a population, and is likely to correspond to what that population considers as "common colours". Some support for this hypothesis can be seen in the fact that the Macquarie Thesaurus [10] lists only 7 principal categories of colour. Experiment: Name 10 colours. Compare your list with the hues in table 1. The Qualification colour model can be implemented by quantising the hue value of a given HSL colour into the small set of base colour names and then prepending an adjective signifying the saturation value and another signifying the luminance. Note that either or both of these adjectives may be omitted, usually indicating mid-range values. Appendix A gives a complete grammar for this process. The order of prepending adjectives has semantic significance. The adjective closer to the colour name appears to bind more strongly and exert a more fundamental modification to the hue. However, the adjective that is read first strongly biases subsequent semantic analysis. Thus "pale bright blue" may be interpreted very differently from "bright pale blue". A third complicating factor is that in general, intensity adjectives seem to exert more influence than saturation adjectives, regardless of position. This is perhaps because the concept of luminance is more directly perceived than that of saturation. The subtleties of adjective order in colour semantics are noted here, but not further investigated in this paper. For the purposes of this experiment, 14 base colours where chosen. Two intensity modifiers were selected and three saturations modifiers. Where saturation or luminance was less than 0.1 or luminance greater that 0.9, all hues where classified as some shade of grey, according to luminance. The resultant colour space consists of 141 distinct colour names. Note that this implementation of CNS differs from Berk et al. in that it uses one fewer saturation and luminance modifiers. Furthermore, so as to differentiate the model from the Sensory model below, these adjectives were chosen so as to be free of obvious emotional connotation. Table 1 summarises the model. An asterisk indicates a special case (grey). A dash indicates a value range for which no modifying adjective is required. Table 1: Qualification colour model Value (of S, L, or H) Saturation adjective Luminance adjective Hue base name < 0.1 * * orange/brown 0.1 0.2 pale dark yellow 0.2 0.3 pale dark green 0.3 0.4 pale aqua 0.4 0.5 pastel aquamarine 0.5 0.6 pastel blue 0.6 0.7 violet 0.7 0.8 purple 0.8 0.9 bright magenta > 0.9 pure * red 3. Sensory model The Sensory model is in some respects a hybrid of the Comparison and Qualification models. The base colour names are the same as those of the Qualification model, but the qualifying adjectives are drawn from a set similar in nature to the colours in the Comparison colour space. The Sensory model hypothesises that people rarely combine more than one adjective when qualifying a colour and that the adjectives they choose tend to be drawn more frequently from sensory or emotional experience than from a strict table of luminance/saturation modifiers. As a result, colour modifiers tend to be more lyrical and analogous than functional. Sensory colour model (examples) HSL triple Colour name [ 0.11 , 0.27 , 0.96 ] pallid yellow [ 0.40 , 0.67 , 0.55 ] dusky green [ 0.58 , 0.22 , 0.56 ] washed-out blue [ 0.62 , 0.64 , 0.67 ] dusky blue [ 0.99 , 0.83 , 0.89 ] soft red Hence, though it may be strictly accurate, people tend not to describe a colour as "dark pale blue" and may even consider this a contradiction. This suggests that the adjectives with which people qualify colour descriptions may be drawn from a single, non-uniform, multidimensional continuum, rather than two orthogonal one-dimensional adjectival spaces. In particular the Sensory model quantises the luminance/saturation plane into a collection of discrete regions, each with a characteristic adjective. These characteristic adjectives are typically epithets transferred from the senses of touch or smell, or from the realm of emotion. Table 2: Sensory colour model adjectives Saturation Luminance < 0.1 0.1 0.3 0.3 0.8 > 0.8 < 0.1 * * * * 0.1 0.4 * gloomy washed-out pallid 0.4 0.6 * murky dull hazy 0.6 0.9 * earthy dusky soft > 0.9 * sombre subdued clear As with the Qualification model, special cases exist when luminance is very low (black) or saturation is low (shades of grey). In these cases (indicated in Table 2 by an asterisk) the adjective applied depends on the hue. Greys which are slightly red or yellow in hue are labelled "warm"; greys which tend towards the blue/green are labelled "cool". Experimental procedure The goal of the research was to determine which, if any, of the three proposed colour naming models best reflected the way individuals name colours. To simplify analysis, the response space was restricted by prompting for selection between fixed alternatives, rather than prompting for a general response. Hence this experiment tested only recognition, not recall of colour names. The subject population chosen was a random sample of 248 people attending the 1991 Monash University Open Day. Participation was informed and voluntary, with testing conducted by unsupervised interaction with carefully designed software. Precautions were taken however to ensure that very young participants were appropriately supervised. Subjects were asked to view a sequence of ten randomly generated colours, projected on a black background. The colours where displayed on a 24-bit RGB screen as a rectangle approximately 35cm by 20cm. As each colour was displayed, subjects were offered a list of three colour names, generated from the HSL values of the colour using each of the three models. The order in which the colour names were listed was randomised to avoid systematic bias. Subjects were asked to select the name which "best" described the colour they were viewing. Subjects preferences were then classified into one of ten categories. Where 70% or more of a subject's responses corresponded to one particular model, the subject was classified as responding strongly to that model (StrongC, StrongQ or StrongS). Where 80% or more of the subject's responses were evenly divided between two models, the subject was classified as responding strongly to those two models (StrongCQ, StrongCS or StrongQS). Where 50% or more of a subject's responses corresponded to one model and no more than 30% of their other responses corresponded to any other model, the subject was classified as showing a slight response to a particular model (WeakC, WeakQ or WeakS). The remaining subjects where classified as showing no response to any particular model (Neutral). The null hypothesis H0 was that responses would be drawn from a uniform random distribution. That is, within each classification that no preference would be seen for any of the proposed models. Under H0 the expected total percentages of population in each classification are: Expected classification given null hypothesis Classification Expected percentage StrongC, Q or S 6.0% StrongCQ, CS or QS 32.2%