Transcript Slide 1

Colour Language 2:
Explaining Typology
Mike Dowman
Language and Cognition
5 October, 2005
Today’s Lecture
• Kay and McDaniel: Direct
neurophysiological explanation
• Terry Regier et al: Predicting denotations
from foci
• Yendrikhovskij: Colours in the environment
• Evolutionary and Acquisitional
Explanations
• Me: An evolutionary model
Kay and McDaniel (1978)
Red, yellow, green and blue colour
categories could be derived directly from
the outputs of opponent process cells
Degree of membership
in colour category
hue
hue
Opponent Processes
Union of blue and green = blue-green
Intersection of red and yellow = orange
Composite categories can be derived using
fuzzy unions
Purple, pink, brown and grey can be derived as
fuzzy using fuzzy intersections
Problems
• Colour term denotations vary across
languages.
• Denotations and foci aren’t in the same
places as opponent process cells predict.
• Doesn’t explain why some types of colour
term are unattested (e.g. blue-red
composites, yellow-green derived terms
(lime)).
Regier et al (2005)
• Is knowing the location of the prototypes in
the colour space enough to predict the full
denotations of colour words?
• Investigated using a computer model.
• Used CIEL*a*b colour space which
attempts to accurately capture conceptual
distances between colours
Details of Computer Model
• Colour categories are represented as
points in the colour space – each at a
unique hue
• Plus a parameter that controls for category
size
• Size parameter was fit to naming data to
get best result
• Each colour is classified based on the
distance to each focus, and the size of the
categories based on each focus
Results: Berinmo
Berinmo naming data:
Model predictions fit
to data:
Categories centred at red, yellow, green, black and white universal foci
• Explains naming in terms of foci
• But doesn’t explain which foci each language uses
• Doesn’t show that non-attested colour term systems
can’t be represented
Yendrikhovskij (2001)
Can the colours in the environment explain typological
patterns in colour naming?
N.B. Photo from
Tony Belpaeme,
not Yendrikhovskij
Distribution of Colours
Full range of colours:
Those in natural images:
• Colours in natural images mapped to CIE
colour space
• Then clustered (those closest to each other
were grouped together)
• Number of clusters was varied
Yendrikhovskij’s Results
• 11 Clusters
 10 are close to centres’ of English colour terms
 A yellow-green cluster replaces purple
• 7 Clusters
 black, white, red, green, yellow, blue, brown
• 3 Clusters
 black, white, red
Distribution of colours in the environment together
with the properties of the ‘sensorial system’
predict attested colour term systems quite well
Acquisitional and Evolutionary
Explanations
Language
Acquisition
Device
Primary
Linguistic Data
Individual's
Knowledge of
Language
Chomsky’s Conceptualization of Language Acquisition.
Language
Acquisition
Device
Primary
Linguistic
Data
Individual's
Knowledge of
Language
Arena of
Language
Use
Hurford’s Diachronic Spiral
Learnable and Evolvable
Languages
E
F
L
Occurring languages
All of the languages which actually exist in the world will fall
within the intersection of the learnable languages, (L), and those
languages which are preferred as a result of evolutionary
pressures, (F) (Kirby, 1999).
Expression-Induction Models
Models simulate the transmission of language
between agents (artificial people)
• Each agent can learn a language based on
utterances spoken by another agent
• In turn they can speak and so create data from
which another agent can learn
L0
L1
L2
Evolving Colour Categories:
Dowman (2003, 2004)
Can we explain colour term typology in terms of
cultural evolution?
This was the original thesis of Berlin & Kay
(1969).
Small biases in the way we learn or perceive
colour categories could create evolutionary
pressures that, over several generations, result
in only a limited range of languages emerging.
Tony Belpaeme (2002) and Me both have
expression-induction models of colour term
evolution
Hypothesis
Typological patterns observed in colour term
naming are due to irregularities in the
conceptual colour space.
 In particular the irregular spacing of the
unique hues
 and their added salience
Agents’ Conceptual Colour Space
red - 7
orange
purple
yellow - 19
blue - 30
green - 26
The whole colour space is 40 units in size
Learning by Bayesian Inference
• Statistical inference allows the most likely
denotation for colour terms to be
estimated based on some example
colours
• Has no predisposition to believe any type
of colour term is more likely than any other
• Can cope with errors in the data
• Each colour word is learned individually
Learning Colour Word Denotations
from Examples
low probability
hypothesis
high probability
hypothesis
medium probability
hypothesis
Urdu
1
0.8
Nila
Hara
0.6
Banafshai
0.4
Lal
Pila
0.2
0
Hue (red at left to purple at right)
Unique Hues
Agent Communication
Agent 3
Agent 3 thinks Mehi is the
best label for colour 27
Nol: 15, 18, 23
Says: Mehi
Wor: 38, 5, 11
Agent 8
Mehi: 27
remembered
by agent 8
Nol: 11, 14
Mehi: 25, 28, 30, 35
Wor: 3, 12
Mehi: 33
Both agents can see: colour 27
Start
A speaker is chosen.
Evolutionary
Model
A hearer is chosen.
A colour is chosen.
Yes (P=0.001)
Decide whether
speaker will be
creative.
The Speaker makes up a new
word to label the colour.
No (P=0.999)
The speaker says the word which they think is most likely
to be a correct label for the colour based on all the
examples that they have observed so far.
The hearer hears the word, and remembers
the corresponding colour. This example will be
used to determine the word to choose, when it
is the hearer’s turn to be the speaker.
Evolutionary Simulations
• Average lifespan (number of colour
examples remembered) set at:
18, 20, 22, 24, 25, 27, 30, 35, 40, 50, 60, 70,
80, 90, 100, 110 or 120
• 25 simulation runs in each condition
Languages spoken at end analysed
• Only agents over half average lifespan
included
• Only terms for which at least 4 examples
had been remembered were considered
Analyzing the Results
Speakers didn’t have identical languages
 Criteria needed to classify language
spoken in each simulation
• For each agent, terms classified as red,
yellow, green, blue, purple, orange, lime,
turquoise or a composite (e.g. blue-green)
• Terms must be known by most adults
• Classification favoured by the most agents
chosen
Example: One Emergent Language
Denotations of Basic Color Terms for all Adults in a Community
Each row is one agent
Each column is a hue
Boxes mark unique hues
Typological Results
25
20
WCS
Simulations
15
10
G-B-R
Y-G-B
R-Y-G
B-R
G-B
Y-G
R-Y
Blue
Green
0
Yellow
5
Red
Percent of terms of this type
30
Type of colour term
Percentage of Color Terms of each type in the
Simulations and the World Color Survey
Derived Terms
•
•
•
•
80 purple terms
20 orange terms
0 turquoise terms
4 lime terms
Divergence from Trajectories
• 1 Blue-Red term
• 1 Red-Yellow-Green term
• 3 Green-Blue-Red terms
Most emergent systems fitted trajectories:
• 340 languages fitted trajectories
• 9 contained unattested color terms
• 35 had no consistent name for a unique hue
• 37 had an extra term
30
25
20
WCS
15
No Unique
Hues
Unique Hues
10
Type of colour term
R-Y-G-B
Y-G-B
G-B-R
R-Y-G
B-R
G-B
Y-G
Blue
R-Y
Green
0
Yellow
5
Red
Percent of terms of this type
Does Increased Salience of
Unique Hues Matter?
Unique Hues Create More Regular
Colour Term Systems
•
•
•
•
644 purple terms
374 orange terms
118 lime terms
16 turquoise terms
Only 87 of 415 emergent systems fits
trajectories
How Reliable is WCS Data?
Would a model that more closely replicated
the WCS data be a better model?
• Field linguists tend to suggest that colours
are much more messy than Kay et al
suggest
• WCS is only a sample – not a gold
standard
• Is data massaged to fit theories?
Summary
• Typological patterns in colour term systems
cross-linguistically can be explained in terms of
uneven conceptual spacing of the unique hues.
• The typological patterns are emergent properties
of the cultural evolution of colour term systems
over time.
• The evolutionary approach readily
accommodates exceptional languages.
• Environmental and/or cultural pressures
probably also influence emergent colour term
systems.
References
Belpaeme, Tony (2002). Factors influencing the origins of color
categories. PhD Thesis, Artificial Intelligence Lab, Vrije Universiteit
Brussel.
Berlin, B. & Kay, P. (1969). Basic Color Terms. Berkeley: University of
California Press.
Dowman, M. (2003). Explaining Color Term Typology as the Product of
Cultural Evolution using a Bayesian Multi-agent Model. In R.
Alterman and D. Kirsh (Eds.) Proceedings of the 25th Annual
Meeting of the Cognitive Science Society. Mahwah, N.J.: Lawrence
Erlbaum Associates.
Dowman, M. (2004). Colour Terms, Syntax and Bayes: Modelling
Acquisition and Evolution. Ph.D. Thesis, University of Sydney.
Hurford, J. R. (1987). Language and Number The Emergence of a
Cognitive System. New York, NY: Basil Blackwell.
Kirby, S. (1999). Function Selection and Innateness: The Emergence of
Language Universals. Oxford: Oxford University Press.
Kay, P. & McDaniel, K. (1978). The Linguistic Significance of the
Meanings of Basic Color Terms. Language, 54 (3): 610-646.
Regier, T. Kay, P. and Cook, R. S. (2005). Universal Foci and Varying
Boundaries in Linguistic Color Categories. In B. G. Bara, L. Barsalou
and M. Bucciarelli (Eds.), Proceedings of the XXVII Annual
Conference of the Cognitive Science Society. Mahwah, New Jersey:
Lawrence Erlbaum Associates.
Yendrikhovskij, S. N. (2001). Computing Color Categories from
Statistics of Natural Images, Journal of Imaging Science and
Technology, 45(5).
Discussion Questions for Tomorrow
• Is colour term typology best explained in terms of
neurophysiology, the environment, cultural practices, or
some other factor?
• What evidence is there for innate biases concerning
colour terms?
• Is colour term evolution really as predictable as Berlin
and Kay’s implicational hierarchy suggests?
• Is it really possible to separate basic from non-basic
colour terms objectively? (Think about English and any
other languages you know.)
• Is colour term typology best explained ontogenetically or
diachronically?