learning word meanings. concept learning review simple associations not enough goal direction /...
TRANSCRIPT
Learning word meanings
Concept learning review
Simple associations not enough• Goal direction / determining tendency• Essences for some types of concept
(“natural kinds”)• Defining features present early for some
concepts (robber)• Characteristic defining for others
(uncle)
Concept learning review ctd
Concept of race
Interaction of universal / innate part with social learning
A “developmental” approach
Concept learning review ctd
But:
• Simple associationism illuminates asymmetric category learning
• Its failures highlight what remains to be explained
• Its limitations don’t mean we can’t model concept learning
Overview of lecture
A. The computational problem
B. Constraints that might help
C. Summary
A. The computational problem
1. Quine’s rabbit
2. Searching a concept space
3. Winston’s arch
Gavagai
Inductive concept learning
(eats-meat fluffy small red) +(eats-meat fluffy big red) -(eats-fruit fluffy small red) -(eats-fruit smooth small red) -(eats-meat fluffy small red) +
What's the concept?(eats-meat fluffy small)
Can a concept like this be learned automatically?
A search problem
For a given number of attributes, a space can be defined of possible concepts
()
(eats-meat) (eats-fruit) (fluffy) (smooth) ...
(eats-meat fluffy) (eats-meat smooth) (eats-fruit fluffy)...
(eats-meat fluffy small)(eats-meat smooth small) ...
etc.
Operators: generalisation & specialisation
Generalisation
Cover more examples:
drop an attribute from a concept
First positive case initialises concept:
(eats-meat fluffy small red) +
(eats-meat fluffy small blue) +
Generalise: (eats-meat fluffy small)
This is a 'move in concept space'
Specialisation
Cover fewer examples:
add an attribute to a concept
(eats-meat fluffy) +
(eats-meat fluffy) -
Specialise: try (eats-meat fluffy small)
… a 'move in concept space'
Winston's arch learnersupports(block1, obj) support(block2, obj) flat(obj) roof(obj)
Ex 1 [+]
Ex 2 [+] supports(block1, obj) support(block2, obj) roof(obj)
DEFINE
GENERALISE
supports(block1, obj) support(block2, obj) roof(obj) not touch(block1, block2)
SPECIALISE
Ex 3 [-]
B. Constraints that might help
1. General expectations
2. Cognitive constraints
3. Language form (syntax) constraints
4. Pragmatic constraints
5. World knowledge
Balaban & Waxman (1997) 9 month old childrenprediction - if child forms category while viewing instance:
1. they get bored (habituate)2. they'll show a novelty preference
is the effect greater with naming?9 rabbits then a pig and a rabbit
More children showed pig preference (sig.)with words than tones accompanying
Waxman & Markow (1995)novelty preference method
12-13 mths - N or Adj (novel word), or no label
Train: 4 instances (eg. 4 animals)
Test: choice of new instance, or non-member
This one is an X Novelty preference
This one is X-ish Novelty preference
Look at this No novelty preference
Waxman & Markow ctd
Words prompt (very young) children to form conceptsA general expectation about word forms
The infants didn’t differentiate between the noun and adjective form
However:Children with a high vocabulary
facilitated superordinate but not basic level category formation
Children with a low vocabularyneither clearly assisted
Booth & Waxman (2002)
Stages 1 and 2: Training
Stage 1 Familiarisation
4 novel objects with characteristic shape & colourThis one is a dax, and this one,…
Look what I can do with this one… [demo]
Look at this one…
Stage 2 Contrast
Booth & Waxman ctd
Stage 3: Generalisation
Forced choice between a new instance and a non-memberCan you find me another one of these?
At 14 mths, demo of function helps
- because it focuses child on a relevant subset of properties
Cognitive constraints
• Perceptual constraints eg. shape
• Constraints can be learned
• Ontological constraint
• Taxonomic constraint
• Mutual exclusivity
Landau, Smith & Jones (1988)
YES NO
Is this a Dax?
YES NO
Does this one match?
Jones, Smith & Landau (1991)Trained example
.50 .53
.76
.82
.48
.80
Soja, Carey, & Spelke (1991)
2 yrs
Novel object introduced, described, and handled
My blicket, this blicket
Then a forced choice: point to the blicket
Soja et al. ctd
Object learned:another same shape different stuff
or three little chunks same stuff
Substance learned:another pile or slick the same
shape, but different stuffor three blobs the same stuff
Soja et al. ctdIf just ask to choose (no trained item, no word)
responses were at chance
another same shape different stuff= three little chunks same stuff
and another pile or slick the same shape,
but different stuff= three blobs the same stuff
Soja et al. ctd
By 2 years
Children know about the distinction between objects and substances
And they use it to organise the generalisation of word meanings
Colunga & Smith (2003)
Previously(Soja, Carey, & Spelke, 1991, Cognition, 38, pp179-211)
Children aged 24/30 monthssolid objects same shapenon-solid objects same materialBut not at 18 months…
Hypothesis: learn this pattern by associationFirst 300 wordsMost denote solid objects, objects that have a consistent shape… and non-solid mostly denote substances i.e. child learns to apply this mapping pattern from associations
present in first words learned
Colunga & Smith (2003)Output units [words]
Hidden units […]
Inputs [shape] [substance] [solid, not s.]
Train: ball [ball-shape] [random] [ 1 0 ]
Test: novel shapes/materials
Prediction - hidden unit activation patterns ("representations") will be similar for
non-solid / same material or
solid / same shape
Colunga & Smith (2003)Prediction - hidden unit activation patterns ("representations") will be similar when
non-solid and same material
solid and same shape
Testing
Forced choice Pick shape
non-solid / same material 30%
solid / same shape 55%
Markman & Hutchinson (1984)
Taxonomic constraint
words refer to whole objects; of same type
3-4 year old children
Target picture eg. poodle
Test pictures eg. alsation or dog foodGive the puppet the one that’s the same.
without label - prefer thematic
with label - prefer taxonomic
Markman & Wachtel (1988)
mutual exclusivity constraint
- two words don’t mean the same thing
Expt 1 (3 years old)
Offer child choice of objects, one unfamiliar. Familiar object already has a name.
Give me a merk
Children tend to choose the novel object
Markman & Wachtel (1988)Expt 2 to check for response bias
Present one object (with a salient part)FAMILIAR fish (fin)UNFAMILIAR microscope (platform)Which is the fripe, the whole thing or just this part?
- What predictions do the constraints make?whole object constraint?mutual exclusivity?
20% chose part for unfamiliar object57% chose part for familiar object
Syntactic constraints
• General expectation differentiates into more specific, syntactically driven, expectations
• Soja
• Language specificity
Syntax – a very brief intro!
Word order indicates relationships among event participantsThe boy kicked the dog
Part of speech is indicated by word order function words, and morphologyThe boyboy function word (= closed class word)
kickkicked morphology (changes word shape)
Syntax – brief intro ctdWord order indicates relationships among event participants
Part of speech is indicated by word order, function words, and morphology
In some languages, morphology can do nearly all the work, and word order matters less (eg. Latin)
Waxman & Booth (2001; 2003)1. Training on 4 purple animals, presented in 2 pairs (same colour, same category)
2. Contrast example orange carrot
3. Then test generalisation
11 mths 14 mths
Nouns
Category [new animal, purple; or purple plate] 0.57 0.68
Property [new animal, purple; or new animal, blue] 0.55 0.44
Adjectives
Category [new animal, purple; or purple plate] 0.59 0.50
Property [new animal, purple; or new animal, blue] 0.58 0.52
No word
Category [new animal, purple; or purple plate] 0.46 -
Property [new animal, purple; or new animal, blue] 0.49 -
Soja, Carey, & Spelke (1991)
2 yrs
Novel object introduced, described, and handled
My blicket, this blicket
Then a forced choice: point to the blicket
Soja et al. ctd
Object learned:another same shape different stuff
or three little chunks same stuff
Substance learned:another pile or slick the same
shape, but different stuffor three blobs the same stuff
Soja et al. ctdIf just ask to choose (no trained item, no word)
responses were at chance
another same shape different stuff= three little chunks same stuff
and another pile or slick the same shape,
but different stuff= three blobs the same stuff
Soja et al. ctd
By 2 years
Children know about the distinction between objects and substances
And they use it to organise the generalisation of word meanings
Soja et al. (1991)
If the learned object was introduced with selective syntax
a blicket
some blicket
… it made no difference
Soja (1992)
2 and 2.5 year olds who had mastered mass-count syntax in their speaking
Were partly sensitive to syntax in word learning
GENERALISES TO
some [substance] substance
a [substance] bounded pile
Language specificityEnglish, Spanish - plural marks nounEnglish - mass/count distinction
draws attention to shape
Korean - classifier language
Experiment (3-5 years; n = 16)novel word applied to an object: "fep", a magnet
choice: cube of same substancewood block same shape
English, Spanish - prefer shape similarKorean - prefer substance
Language specificity ctd
But, classifiers highlight shape:
Empitsu o gohon kudasi
pencil five long thin given
Yonpil tasot caru
pencil five long thin
Pragmatic influencesPrinciple of contrastClark (1993)
- every difference of form marks difference in meaning
- economical for learning- a pragmatic principle
-- used once understand speaker is intentional
For Clark, contrast means any difference in meaning (including connotation, register & dialect).
Identity of reference is not sufficient
Tomasello & Barton (1994) DevPsych 30 639-650
2 years "Let's go find the toma"
look in one of buckets (5)
Either find it straight away
or
first find and reject two
("oh no", scowl, put back)
then find the right thing
Akhtar & Tomasello (1996) 2 yearsSimilar expt but one (distinctively shaped)
bucket is shut and can't be opened
Pre-play, so that child is familiar with the objects in each bucket (no naming)
Put them backAdult - "Now, let's find the toma!"Adult expresses disappointment at no access,
but plays with other objectsLearned equally well whether no access or did
retrieve
Role of world knowledge Schank, Collins & Hunter (1986)
• Hijackings cubageneralisation? cuba?
• Hijacking libyasyntactically, do what?drop destination as a dimension?or generalise feature content?
e.g. warm country?
Target concept - a model of how terrorists select destinationsWhich are relevant features has to be worked out
often not perceptually available
C. Summary
1. Quine’s rabbit & the problem
2. Constraints guide search of the space
3. A variety of factors influence learning word meanings