Simple interactions between shortterm memory and long-term
knowledge explain large amounts of
developmental phenomena
Gary Jones
Nottingham Trent University (UK)
Gary.Jones@NTU.ac.uk
In the knowledge is the power
• No serious developmental researcher would
argue against knowledge being a key factor in
developmental change
• BUT the knowledge a child brings to bear in a
particular domain is usually hard to
approximate and define (e.g. difficult to
ascertain the exposure to the task)
• Language is perhaps the only domain where one
can hypothesise accrued knowledge by
approximating linguistic exposure
In the knowledge is the power
• What (basic) linguistic knowledge do children
learn?
• Phoneme sequences (Saffran, 2001 amongst
others)
▫ golatutibudodaropitibudogolatudaropitibudo….
• Word sequences (Fernald et al., 1998)
doggy
where’s the doggy
Computational model (of associative
language learning)
• Phonological knowledge (phoneme sequences)
and lexical knowledge (words, phrases) are
learnt from large-scale samples of child-directed
phonological-based language input (maternal
utterances, literature aimed at 4-5 year olds)
• Computational model constrains the input and
learns from it
Digit and ‘other’ spans in children
STM + (3 x LTM)?
STM + (2 x LTM)?
STM + (1 x LTM)?
Chunking and digit span
• We are repeatedly exposed to random/pseudorandom sequences of digits
▫ Phone numbers, room numbers, dates, times
• Some of these will become chunks e.g. 2013
• Even if we chunk only some of the digit
sequences, it is still likely they will help in the
recall of a random digit sequence
Chunking and digit span
• Chunking requires repeated exposure to stimuli
• In order to estimate the chunks a child may
learn, need to use a domain where exposure to
the stimuli can be reasonably estimated
▫ Language: prior to beginning to read, young
children’s (2-5 years of age) language input can be
estimated from the language spoken by primary
caregivers and from children’s story books
A chunking account of spoken language
acquisition
• Gradually chunk the input into increasingly
larger chunks
▫
▫
▫
▫
/
/
/
(“where’s Mummy?”)
A chunking account of spoken language
acquisition
• Fixed STM capacity that can hold 4 chunks of
information
▫
▫
▫
▫
/
/
/
= 8 chunks
= 4 chunks
= 2 chunks
= 1 chunk
The chunking account (same span tests
as per the 8 and 10 year old children)
Additional support for chunking
account: Nonword repetition
Additional support for chunking
account: Nonword repetition
Additional support for chunking
account: Nonword repetition
Additional support for chunking
account: Nonword repetition
Testing the chunking account of digit
span
• Present participants with digit sequences that
they are likely to have been repeatedly exposed
to
• For example, testing participants in lab 425 with
a digit span containing the (familiar) digit
sequence 4, 2, 5 compared to an unfamiliar digit
sequence such as 5, 2, 4
Testing the chunking account of digit
span
% familiar digit
% unfamiliar digit
chunks recalled
chunks recalled
57.86 (23.09)
34.38 (26.70)
t(43) = 5.78, p < .001
Testing the chunking account of digit
span
• But: we are unable to predict what digit
sequences are meaningful to participants (except
for sequences like 1, 2, 3; 5, 7, 9 etc.)
• We can therefore only estimate the benefit by
averaging performance across digit sequences
compared to other sequences
Testing the chunking account of digit
span
• Presenting digit lists, word lists, and mixed lists
• For mixed lists, change odd numbers in digit
span lists or change even numbers
▫ Even numbers: might, eight, two, six, had
▫ Odd numbers: three, earth, best, box, five
• Compare likelihood of accuracy for digit
sequences versus word sequences in mixed lists
Testing the chunking account of digit
span
p = .003
p < .001
12
Span size
10
8
6
4
2
0
Digit span
Mixed span
Span test
Word span
Testing the chunking account of digit
span
Mixed lists:
Mixed lists:
% digit pairs recalled
% word pairs recalled
.60 (.30)
.50 (.30)
t(81) = 2.00, p = .049
Testing the chunking account of digit
span
• For a digit sequence containing 5 digits:
▫ Accuracy = .6 x .6 x .6 x .6 = .13 (13% chance of
accuracy)
• For a word sequence containing 5 words:
▫ Accuracy = .5 x .5 x .5 x .5 = .06 (6% chance of
accuracy)
• Twice as likely to correctly recall a digit sequence
Summary
• Performance on the digit span task is superior to
performance on other types of span task because
sequences of digits occur in our environment more
often than sequences of other stimuli
▫ We therefore learn a greater number of chunked digit
sequences than (for example) sequences of random
words
• Digit span is not a valid measure of STM capacity
▫ Arguably, neither are many other measures for similar
reasons
Acknowledgements
Daniel Freudenthal
Fernand Gobet
Helen Hearn
Julian M. Pine
The Leverhulme Trust