See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/23224085

Numerical thought with and without words: Evidence from Indigenous

Australian children

Article in Proceedings of the National Academy of Sciences · October 2008

DOI: 10.1073/pnas.0806045105 · Source: PubMed

CITATIONS READS
126 645

4 authors:

Brian Butterworth Robert Reeve

University College London University of Melbourne
223 PUBLICATIONS 17,126 CITATIONS 77 PUBLICATIONS 3,450 CITATIONS

SEE PROFILE SEE PROFILE

Fiona Reynolds Delyth Lloyd

University of Melbourne Department of Health and Human Services
6 PUBLICATIONS 380 CITATIONS 16 PUBLICATIONS 863 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Robert Reeve on 23 May 2014.

The user has requested enhancement of the downloaded file.

Numerical thought with and without words: Evidence
from indigenous Australian children
Brian Butterworth*†, Robert Reeve†‡, Fiona Reynolds‡, and Delyth Lloyd*
*Institute of Cognitive Neuroscience, University College London, 17 Queen Square, London WC1N 3AR, United Kingdom; and ‡Department of Psychology,
University of Melbourne, Melbourne VIC 3010, Australia

Communicated by Rochel Gelman, Rutgers, The State University of New Jersey, Piscataway, NJ, July 8, 2008 (received for review April 10, 2008)

Are thoughts impossible without the words to express them? It has In the study we report here, we contrasted three languages:
been claimed that this is the case for thoughts about numbers: Warlpiri, Anindilyakwa, and monolingual English. Warlpiri is a
Children cannot have the concept of exact numbers until they classifier language spoken in the Central Desert north and west
know the words for them, and adults in cultures whose languages of Alice Springs, Northern Territory (NT). It has three generic
lack a counting vocabulary similarly cannot possess these concepts. types of number words: singular, dual plural, and greater than
Here, using classical methods of developmental psychology, we dual plural. Anindilyakwa, another classifier language, is spoken
show that children who are monolingual speakers of two Austra- on Groote Eylandt, NT, in the Gulf of Carpentaria. It has four
lian languages with very restricted number vocabularies possess possible number categories: singular, dual, trial (which may in
the same numerical concepts as a comparable group of English- practice include four), and plural (more than three) (12). There
speaking indigenous Australian children. are also loan words used as number names for 1, 2, 3, 4, 5, 10,
15, and 20, but these appear to be used only in certain contexts
cognitive development 兩 linguistic determinism 兩 mathematical cognition 兩 and children do not know them (13). Neither language has
number concepts ordinals equivalent to ‘‘first,’’ ‘‘second,’’ ‘‘third’’; both have
quantifiers similar to ‘‘few’’ and ‘‘many’’ (12). [For further details
about both languages, see supporting information (SI) Text]. We
A strong form of the hypothesis that language determines
thought (the Sapir-Whorf hypothesis) (1) has been revived
for the domain of numbers (2, 3) on the basis of studies of
also tested monolingual English-speakers in Melbourne, Aus-
tralia at a school for indigenous children. See Fig. 1 for NT
locations.
Amazonian cultures with languages that lack counting words (4,
5), even though the hypothesis has been largely abandoned Results
elsewhere (6,7). It is argued that a vocabulary of counting words

PSYCHOLOGY
We tested 45 children aged 4 to 7 years old: 20 Warlpiri-speaking
is necessary for a person to possess the concepts of exactly four, children, 12 Anindilyakwa-speaking children, and 13 English-
exactly five, and so forth (2, 3). Without the vocabulary, only speaking children from Melbourne. Approximately half the NT
primitive, approximate numerical values are possible (2, 8). It is children were 4 to 5 years old and half were 6 to 7 years old. We
proposed that counting words modify two innate core systems of used four enumeration tasks to evaluate numerosity understand-
knowledge with numerical content (2,8): parallel individuation ing: memory for number of counters, cross-modal matching of
of objects, which enables the representation of exact numerosi- discrete sounds and counters, nonverbal exact addition, and
ties up to three, and analogue magnitudes that represent ap- sharing play-dough disks that could be partitioned by the child
proximate numerosities of more than three. According to one (see Methods and SI Text)
version, children learn to associate the words ‘‘one,’’ ‘‘two,’’ and
‘‘three’’ with the state of the parallel individuation system, and Memory for Number of Counters. No language effects were found
generalize from this that other number words also denote exact (F ⬍ 1) (see Fig. 2A). Children were more accurate recalling
numerosities (2). On another account, children make use of the small, compared to large numerosities [F (1, 24) ⫽ 16.05, P ⬍
fact that they have already associated larger number terms with 0.001], and older NT children recalled more than their younger
approximate numerosities (9), and refine their sense of, for peers [F (1, 28) ⫽ 16.30, P ⬍ 0.001]. No other effects were found.
example, approximately fiveness into exactly fiveness (5,10). It
follows from these accounts that a child raised without linguistic Cross-Modal Matching. No language effects were observed (F ⬍1
means for representing increasing exact numerosities will not be Fig. 2B). Young children in all locations were more accurate at
able to develop concepts of the natural numbers, each denoting cross-modal matching small compared to large numerosities [F
an exact numerosity with a unique successor. (1, 18) ⫽ 14.82, P ⬍ 0.001]. Older NT children were more
Evidence for the strong form of Whorf’s hypothesis comes accurate than their younger peers [F (1, 25) ⫽ 5.41, P ⬍ 0.03].
from two studies of the numerical abilities of speakers whose No interactions were observed.
languages have restricted number-word vocabularies. Adult
speakers of Pirahã, an Amazonian language that contains words Nonverbal Addition. Although Melbourne children solved fewer
for just ‘‘one,’’ ‘‘two,’’ ‘‘few,’’ and ‘‘many,’’ have difficulty putting problems correctly than their NT peers, the difference was not
small sets of objects in one-to-one correspondence, and fail in a significant (P ⬎ 0.1) (Fig. 2C). Children solved more simple than
task working out the consequence of adding to, or subtracting
one item from, a small set of objects (4). Another Amazonian Author contributions: B.B. and R.R. designed research; F.R. and D.L. performed research;
group, the Mundurukú, whose language contains words for exact B.B. and R.R. analyzed data; and B.B. and R.R. wrote the paper.
numbers to about three and approximate numbers to about five, The authors declare no conflict of interest.
perform comparably with French adult controls on tasks involv- Freely available online through the PNAS open access option.
ing approximate numerosities, but are much worse than controls †To whom correspondence may be addressed. b.butterworth@ucl.ac.uk or r.reeve@
on simple exact subtraction (5). Both Amazonian groups are unimelb.edu.au
hunter-gatherers whose lifestyles differ from our own in many This article contains supporting information online at www.pnas.org/cgi/content/full/
ways, but the factor held responsible for the difference on 0806045105/DCSupplemental.
number tasks is their limited vocabulary of number words (11). © 2008 by The National Academy of Sciences of the USA

www.pnas.org兾cgi兾doi兾10.1073兾pnas.0806045105 PNAS 兩 September 2, 2008 兩 vol. 105 兩 no. 35 兩 13179 –13184

 Fig. 1. Location map of Willowra and Angurugu.

difficult problems [F (1, 17) ⫽ 15.34, P ⬍ 0.001 for young 2D). Children were less successful sharing seven and ten items:
children in all locations; F (1, 19) ⫽ 13.96, P ⬍ 0.001 for the two only some older NT children successfully partitioned and dis-
NT groups]. Older NT children solved more problems than their tributed the spare play dough disk appropriately. [7 items ␹2 (2,
younger peers [F (1, 19) ⫽ 13.38, P ⬍ 0.001]. A significant n ⫽ 33) ⫽ 8.07; 10 items ␹2 (2, n ⫽ 33) ⫽ 6.12, p’s ⬍ .05].
interaction between set size and age [F (1, 19) ⫽ 4.67, P ⬍ 0.05] To understand how exact number was varied as a function of
was attributable to the younger children’s relative inability to individual differences, hierarchical cluster analyses (using
solve the more difficult problems (P ⬍ 0.05). Ward’s method) were performed on each measure to identify
task-related low, medium, and high performances. Tables 1 and
Sharing. Almost all children were able to share six or nine items 2 show that children with the highest level of competence were
among the three bears, typically using one-to-one dealing (Fig. from all sites.

13180 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0806045105 Butterworth et al.

 4- to 5-year-olds 4- to 5-year-olds
A 100 6- to 7-year-olds B 100 6- to 7-year-olds
90 90
Unhatched = small n sets Unhatched = small n sets
80 Hatched = large n sets 80 Hatched = large n sets
70 70
Percentage Correct

60 60

50 50

40 40

30 30

20 20

10 10

0 0
Eng li sh Walpiri Anindilyakwa Eng li sh Walpiri Anindilyakwa

Language Language

C 100 4- to 5-year-olds D 100 English

90 6- to 7-year-olds 90 Walpiri
Unhatched = small n sets Anindilyakwa
80 80
Hatched = large n sets
70 70
Percentage Correct

60 60

50 50

40 40

30 30

20 20

10 10

0 0
Eng li sh Walpiri Anindilyakwa 6 9 7 10
Language Set Size

PSYCHOLOGY
Fig. 2. Percentage of correct responses as a function of age, set size and language for the memory for counters (A), cross-modal matching (B), nonverbal addition
(C), and sharing (D) tasks.

There were significant relationships between performance linear trend for each language group for all tasks, with r2 values
and age on the memory for counters, nonverbal addition, and between 0.80 and 0.99 and no observed discontinuities between
sharing measures. Older children were over-represented in the small and large numbers (see SI Text for linear trends.) MANO-
high performance groups [memory for counters: ␹2 (2, n ⫽ 45) ⫽ VAs, adjusted for age, confirmed that there was no difference
12.82, P ⬍ 0.002; nonverbal addition: ␹2 (2, n ⫽ 32) ⫽ 12.77, P ⬍ between groups. The scalar variability of responses (coefficients
0.002; sharing: ␹2 (2, n ⫽ 33) ⫽ 6.88, P ⬍ 0.03]. of variation) was not significantly different from zero for the
We also analyzed the relationship between responses and tasks in all language groups (see SI Text). This is consistent with
targets to determine whether there was a discontinuity between the use of nonverbal enumeration, but not verbal counting, for
small (ⱕ4) and large numbers (⬎4) (Fig. 3 A–C). There was a all numerosities in these tasks (14).
In this study, no language effects were observed. Neither the
Warlpiri-speaking nor Anindilyakwa-speaking children per-
Table 1. Number of children assigned to low, medium, and high formed worse than the English-speaking children on any task.
competence cluster groups as a function of task and language Failure to find performance differences was not because of the
Task and language Low Medium High

Memory for counters Table 2. Number of children assigned to low, medium, and high
English 8 2 3 competence cluster groups as a function of task and age
Warlpiri 7 6 7
Task and age Low Medium High
Anindiyakwa 4 4 4
Cross-modal matching Memory for counters
English 3 5 1 4 and 5 year olds 17 7 3
Warlpiri 2 8 8 6 and 7 year olds 2 5 11
Anindiyakwa 1 6 3 Cross-modal matching
Nonverbal addition 4 and 5 year olds 4 13 5
English 5 3 1 6 and 7 year olds 2 6 7
Warlpiri 5 1 7 Nonverbal addition
Anindiyakwa 2 2 6 4 and 5 year olds 11 6 5
Sharing continuous quantity 6 and 7 year olds 1 9
English 3 7 Sharing continuous quantity
Warlpiri 1 10 4 4 and 5 year olds 3 15 1
Anindiyakwa 5 3 6 and 7 year olds 1 7 6

Butterworth et al. PNAS 兩 September 2, 2008 兩 vol. 105 兩 no. 35 兩 13181

 A 12
English
B 12
English
11 Walpiri 11 Walpiri
Anindilyakwa Anindilyakwa
10 10

Children’s Responses
9
Children’s Responses

8 8

7 7

6 6

5 5

4 4

3 3

2 2

1
. = One Child’s
= One Child’s Answer
Response 1 . = One C=hild
One child’s response
’s Response
0
0
2 3 4 5 6 8 9 1 2 3 4 5 6 7
Set Size
Set Size Set Size

C 12
English
11
Walpiri
10
Anindilyakwa
Children’s Responses

1 . = One Child’s
= OneResponse
child’s response

0
2+1 1+2 3+1 1+3 4+1 1+4 3+3 4+2 5+3

Addition Problems
Fig. 3. Distribution of children’s responses as a function of set size for the memory for counters (A), cross-modal matching (B), and nonverbal addition (C) tasks.

insensitivity of the tests, as one predictor variable significantly Alternative accounts propose that we are born with a capacity to
related to performance: namely, children’s age. If a number represent exact numerosities (17–19), and that using words to
vocabulary were necessary for the development of exact number name exact numerosities is useful but not necessary (11, 20).
concepts, then no NT children should have achieved high levels When children learn to count, they are learning to map from
of numerical competence, yet high levels were reached by both their pre-existing concepts of exact numerosities onto the count-
NT groups in all tasks (see Fig. 4 A to C for response frequency ing word sequence (11, 20). Conceptual development drives the
distributions). Discontinuities in accuracy between small and acquisition of counting words rather than the other way around.
large numbers is held to be the signature of the two core systems
operating without the aid of language (2, 3, 8). Here we showed Methods
that tasks became harder as the number increased, with no In Willowra and Angurugu, bilingual indigenous assistants were trained by an
discontinuities in the linear trend from 1 to 9 objects for all experimenter (D.L. or F.R.) to administer the tasks, and all instructions were
groups (see SI Text). given by a native speaker of Warlpiri or Anindilyakwa. Procedures used to
familiarize NT children and indigenous assistants with tasks and materials are
Methodological differences between this study and the Am-
described in the SI Text. Piloting showed children could easily grasp the
azonian studies may account for the conflicting results. The
purpose of the tasks. The experimenter recorded relevant aspects of task
adults in the Pirahã study (4) may not have understood the tasks performance as they occurred.
(15), and in the Mundurukù study (5) subtraction was the only
exact number task, and this operation is difficult for children 4 Memory for Counters. Identical 24-cm ⫻ 35-cm mats and bowls containing 25
to 5 years of age (16) (see SI Text). counters were placed in front of a child and the experimenter. The experi-
We conclude that the development of enumeration concepts menter took counters from her bowl and placed them on her mat, one at a
does not depend on possession of a number-word vocabulary. time, in preassigned locations. Four seconds after the last item was placed on

13182 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0806045105 Butterworth et al.

 Target Numerosity
A 100
10 0
Target Numerosity
Target 1
Target 1
Target 2
80 Target 2
Target 3 80
Target 3
Target 4
Response Frequency

Target 4

Response Frequency
Target 5
60 Target 5
60
Target 6
Target 6
Target 7
Target 7
40 40

20 20

0 0
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12

Target Numero sity Target Numero sity

B Target Numerosity
100 Target Numerosity
100
Target 1 Target 1

Target 2 Target 2
80
80
Target 3 Target 3
Response Frequency

Target 4 Target 4
60 Response Frequency
Target 5 60 Target 5

Target 6 Target 6

40 Target 7 Target 7
40

20
20

PSYCHOLOGY
0
0
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Target Numerosity Target Numerosity

C Target Numerosity Target Numerosity

100 100
Sum = 3 Sum = 3

Sum = 4 Sum = 4
80 80
Sum = 5 Sum = 5
Response Frequency

Response Frequency

Sum = 6
60 60 Sum = 6

40 40

20 20

0 0
2 3 4 5 6 7 8 2 3 4 5 6 7 8

Target Numerosity Target Numerosity

Fig. 4. Target numerosity frequency graphs for English-speaking (Left) and combined indigenous language groups (Right) for the memory for counters (A),
cross-modal matching (B), and nonverbal addition (C) tasks.

the mat, all items were covered with a cloth and children were asked by the indigenous assistant said, ‘‘Like this? Yes!’’ The experimenter then tapped the
indigenous assistant to ‘‘make your mat like hers.’’ Following three practice blocks three times and placed three counters on her mat—the indigenous
trials in which the experimenter and an indigenous assistant modeled recall assistant said: ‘‘Like this? Yes!’’ The experimenter tapped the blocks three
using one and two counters, children completed 14 memory trials comprising times again, but placed only two counters on the mat. The indigenous assis-
two, three, four, five, six, eight, or nine randomly placed counters. In modeling tant said: ‘‘Like this? No!’’ The experimenter placed a third counter on the mat
recall, counters were placed on the mat without reference to their initial and the indigenous assistant said: ‘‘Like this? Yes!’’ Seven trials, comprising
location. Number and locations of children’s counter recall were recorded. numerosities one to seven, were presented in a random order.

Cross-Modal Matching. The experimenter demonstrated the task by tapping Nonverbal Addition (21). Using materials from the memory task, the experi-
two wooden blocks once and placing a single counter on the mat, while the menter placed one counter on her mat and, after 4 seconds, covered her mat.

Butterworth et al. PNAS 兩 September 2, 2008 兩 vol. 105 兩 no. 35 兩 13183

Next, the experimenter placed another counter beside her mat and, while the that is, whether it was given to one bear or an attempt was made to divide it
child watched, slid the additional counter under the cover and onto her mat. among the three bears.
Children were asked by the indigenous assistant to ‘‘make your mat like hers.’’
Nine trials comprising 2 ⫹ 1, 3 ⫹ 1, 4 ⫹ 1, 1 ⫹ 2, 1 ⫹ 3, 1 ⫹ 4, 3 ⫹ 3, 4 ⫹ 2, and ACKNOWLEDGMENTS. We thank Robert Hoogenraad and Mary Laughren,
5 ⫹ 3 were used. Children’s answers were recorded. experts in Warlpiri language and culture; Barbara Wilson and Julie Waddy,
who provided much-needed assistance and support on Groote Eylandt; Sue
McAvoy, Frank Atkinson, and Gina Atkinson for their help and support at
Sharing. This task assessed the ability to share quantities of play-dough among Willowra; and Nadja Reiter of the Northern Territory Education Department,
three toy bears. Although the play-dough disks comprised equal-sized discrete who helped us overcome many tricky problems. Edith Bavin helped us in the
units (3-cm disks), each disk-unit could be regarded as a continuous quantity early stages of the study, and Judi Humberstone contributed to the statistical
for sharing purposes. Following two practice trials in which children shared analyses. We particularly thank the children and adults of the indigenous
communities in which we worked, who gave their time freely, patiently, and
four disks between two bears (‘‘give these to the bears’’), they completed four
with good humor. We also thank an anonymous reviewer for suggesting that
randomly-ordered trials comprising 6, 9, 7, and 10 disks, which they shared we use Bayesian likelihood ratios to show an absence of difference between
among the three bears. The experimenter recorded the number of disks given numerical response distributions. The research was supported by a grant from
to each bear, the sharing strategies, and the treatment of any remainder disk; the Leverhulme Trust (to B.B. and R.R.).

1. Whorf BL (1956) Language, Thought and Reality. (MIT Press, Cambridge, MA). 12. Bittner M, Hale K (1995) in Quantification in Natural Languages, ed Bach E (Kluwer,
2. Carey S (2004) On the origin of concepts. Daedulus 133:59 – 68. Dordrecht), pp 81–107.
3. Le Corre M, Carey S (2007) One, two, three, four, nothing more: An investigation of the 13. Stokes B (1982) in Language and Culture. Work Papers of SIL-AAB, Series B, ed
conceptual sources of the verbal counting principles. Cognition 105:395– 438. Hargrave S (1982), Vol. 8.
4. Gordon P (2004) Numerical cognition without words: Evidence from Amazonia. Sci- 14. Cordes S, Gelman R, Gallistel CR (2001) Variability signatures distinguish verbal from
ence 306:496 – 499. nonverbal counting for both large and small numbers. Psychonomic Bull Rev 8(4):698 –
5. Pica P, Lemer C, Izard V (2004) Exact and approximate calculation in an Amazonian 707.
indigene group with a reduced number lexicon. Science 306:499 –503. 15. Everett D (2005) Cultural constraints on grammar and cognition in Pirahã. Another
6. Fodor J (1975) The Language of Thought. (Thomas Y. Crowell, New York, NY). look at the design features of human language. Curr Anthropol 46:621– 646.
7. Pinker S (1994) The Language Instinct. (The Penguin Press, London). 16. Starkey P, Gelman R (1982) in Addition and Subtraction: A Cognitive Perspective. eds
8. Feigenson L, Dehaene S, Spelke E (2004) Core systems of number. Trends Cognit Sci Carpenter TP, Moser JM, Romberg TA (LEA, Hillsdale, NJ), pp 99 –116.
8(7):307–314. 17. Butterworth B (1999) The Mathematical Brain. (Macmillan, London).
9. Gilmore CK, McCarthy SE, Spelke ES (2007) Symbolic arithmetic knowledge without 18. Gallistel CR (2007) Commentary on Le Corre & Carey. Cognition 105:439 – 445.
instruction. Nature 447:589 –592. 19. Gelman R, Gallistel CR (1978) The Child’s Understanding of Number. (Harvard Univ
10. Dehaene S, Spelke E, Pinel P, Stanescu R, Tsivkin S (1999) Sources of mathematical Press, Cambridge, MA).
thinking: behavioral and brain-imaging evidence. Science 284:970 –974. 20. Locke J (1690/1961) An Essay Concerning Human Understanding. (JM Dent, London).
11. Gelman R, Butterworth B (2005) Number and language: how are they related? Trends 21. Mix KS, Huttenlocher J, Levine S (1996) Do preschool children recognize auditory-visual
Cognit Sci 9(1):6 –10. numerical correspondences? Child Dev 67:1592–1608.

13184 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0806045105 Butterworth et al.

View publication stats