Transcript ppt

CS 621 Artificial Intelligence
Lecture 35 - 11/11/05
Guest Lecture by
Prof. V Sarma
Human Cognition: The issue of
Language Acquisition
Do you remember learning to...
• tie your shoe laces?
• ride a bicycle?
• to march?
• speak?
You may recall the first three. But not the
fourth – other than anecdotes of baby talk
from parents
And you received explicit instruction and
training to perform the first three but not to
speak your first language(s)
How is this possible?
What does the child do or not do in acquiring
a language?
• Children do not and cannot store all possible
words and sentences since children learn to
understand and produce (an infinite number)
of novel sentences.
• Children learn the “rules” to use their
language creatively even though no one
teaches them these rules (cf. explicit
instruction).
• Children learn the words, sound and
sentence structure of their native
language(s).
Stages of language acquisition1
Language acquisition follows predictable sequences
but the absolute ages varies from child to child
Stage
Babbling
One-word
Two-word
Multiword 1
not
Multiword 2
Age
Description
0;6 - 0;8 repetitive CV patterns
0;9 - 1;6 Single words or word stems
1;6 - 2;0 mini-sentences
2;0 - 2;6 telegraphic" sentences with lexical
functional/grammatical morphemes
2;6 on Grammatical/functional morphemes
Stages of language acquisition2
What do children know at various stages of
development? Is their production a direct
representation of their perception/knowledge?
What is innate and what is learned? We discuss
examples from the grammatical modules of
• Phonology
• Lexicon, word learning
• Morphology
• Syntax
Phonology: Categorization - 1
Phoneme categorization (Eimas, Siqueland,
Jusczyk, and Vigorito 1971):
Issue: Do infants perceive speech sounds categorically?
Subjects: 1 and 4-month-old babies acquiring English
Stimuli: tokens on the /ba/-/pa/ continuum (<25ms /b/)
VOT (msec)=
0
20 40 60 80
Adult perception: [------ba-------][-------pa------]
Babies sucked on pacifiers connected to a computer. First
one syllable was played over and over, then the stimulus
changed.
Pairs of input:
[20, 40],
[60, 80],
[0,20]
Results:
increase
no change no change
Phonology: Categorization - 2
Are these categories learned or innate? Lasky,
Syrdal-Lasky, and Klein (1975)
Subjects: 4-6.5-mo old Guatemalan babies acquiring
Spanish
Stimuli: tokens on the /ba/-/pa/ continuum
[20, 60] voiced-voiceless boundary for English speakers
[-20, 20], boundary between voiced-voiceless stops of
Spanish
[-60, -20] boundary between Thai pre-voiced and voiced
stops
Do the infants’ responses change because Spanish VOT is
different from English VOT? Are they sensitive to
categorization (prevoicing)that is not seen in their
language
Phonology: Categorization - 3
VOT=
-60
-20
20
60
Spanish adults: [-------ba-------][-------pa-------]
English adults: [--------ba-------------][--pa--]
Thai adults:
[--bba--][-------ba-------][--pa--]
Results:
20/60 Voiced-voiceless Increase in
English stops
heartbeat
-60, -20 Prevoiced-voicedIncrease in
Thai stops
heartbeat
(Same results found with Kikuyu and English speaking children)
-20, 20
Voiced-voiceless
CHANGE
Spanish stops
NO
Phonology: Categorization - 4
The effect of language experience on perception
(Werker and Tees 1984)
• 6-8 month-old infants discriminate between
contrasts not made in English (retroflexes in
Hindi) like Hindi speaking adults do but not
English-speaking adults; 95% accuracy
• 8-10 months decline in discrimination ability, 6070% accuracy
• 10-12 months apparent loss of ability, 20%
accuracy
Phonology - Production
Production lags behind perception: ‘Fis’
phenomenon (Berko and Brown 1960)
• Child (pointing at aquarium):
/fis/
Adult:
Huh?
C:
/fis/
A:
Oh, fis.
C:
No fis
A:
Oh! Fish!
C:
Ye fis
Learning word meaning
Innate biases or expectations that constrain the
hypothesis space for learning word meaning
• Whole object bias: Nouns must refer to whole objects.
• Taxonomic bias: Nouns refer to classes of objects and
not to one specific example. But children only hear
specific examples!
• Shape bias: shape plays a more important role in
defining object categories than color, size, material, etc.
• Mutual exclusivity bias: words always have different
meanings (ranked above Whole object bias)
Acquiring morphological rules -1
This was the first evidence that psycholinguists
found that kids are developing a grammar; they
can produce new words that they never heard
before. (Jean Berko 1958)
Test: “Here is a wug. Now there are two of them. There are
two ...?” (Answer, /wugz/)
(English plural rule, with allomorphic variation)
heafs, wugs, luns, tors, cras, tasses, gutches, kashes,
nizzes
Results: Three year olds perform with great accuracy
Acquiring morphological rules -2
English past tense morphemes
• My teacher holded the baby rabbits and we
patted them
• Hey, Horton heared a Who
• I finded the ball
• The alligator goed kerplunk.
Not produced by adults, over-regularization.
Acquiring morphological rules -3
Compound formation: Kiparsky noted that
compounds can be formed from irregular
but not regular plurals:
• men-bashing
• mice-infested
• teethmarks
*gays-bashing
*rats-infested
*clawsmarks
Irregular plurals are stored in the mental
dictionary as roots or stems and can enter
word-formation; compounding takes place
with stored words.
Acquiring morphological rules -4
Gordon 1986
Subjects: 3-5 year olds
“Here is a monster that eats mud. He is called a mudeater.”
“… who likes to eat mice” MICE-EATER (MOUSE-EATER)
“… who likes to eat rats” RAT-EATER
“… who likes to eat purple people” PURPLE-PEOPLEEATER
“… who likes to eat purple lions” PURPLE-LION-EATER
Question: Learning from parents by looking at pluralcompounds?
No plural compounds in motherese (poverty of stimulus)
Automatically distinguish between stored and derived
items.
Acquiring syntax - 1
Do children have syntax in the one-word stage?
(Hirsh-Pasek et al. 1985)
Subjects: 17-mo-olds, one-word stage, using few/no verbs.
They watched two TV's showing familiar characters from
the Sesame Street TV show: Cookie Monster and Big
Bird.
Task: One TV showed CM tickling BB; the other TV showed
BB tickling CM. The babies heard one of these
sentences:
(1) “Cookie Monster is tickling Big Bird.”
(2) “Big Bird is tickling Cookie Monster.”
Results: the babies looked longer at the correct TV scene
Acquiring syntax - 2
(Hirsh-Pasek et al. 1988)
Found the same result for 24-month-olds with
more subtle syntactic cues, and without any
lexical cues:
• “Big Bird is flexing Cookie Monster.” [CAUSE
TO]
• “Big Bird is flexing with Cookie Monster.”
[ALONG WITH]
Thus young kids can use syntactic cues in
comprehension even before they can produce
whole sentences.
Acquiring syntax - 3
Children make certain errors (goed, holded, oxes
etc.) but never certain others are logically
plausible
In all languages it is impossible to move a "whphrase" out of a conjoined noun phrase (as in
2b)
• 1a. John ate eggs with toast
1b. What did John eat eggs with?
• 2a. John ate eggs and toast?
2b. *What did John eat eggs and?
Acquiring syntax - 4
Knowledge of wanna-contraction in English
• 1a
1b
• 2a
2b
• 3a
3b
3c
I want to eat a cookie.
I wanna eat a cookie.
What do you want to eat?
What do you wanna eat?
You want who to eat a cookie?
Who do you want to eat a cookie?
*Who do you wanna eat a cookie.
Generalization: contraction is not possible when
questioning the subject of the subordinate
clause.
Acquiring syntax - 5
An Experiment to test wanna-contraction in
children aged 3-5 years. The experimental
protocol is designed to elicit questions in which
the object of the subordinate clause is
questioned
Exp: The rat looks hungry. I bet he wants to eat something.
Ask
Ratty what he wants.
Child: What do you wanna eat?
Rat: Some cheese would be good.
Acquiring syntax - 6
The next protocol is used to elicit questions about
the subject of the subordinate clause
Exp: There are three guys in this story: Cookie Monster, a
dog, and this baby. One of them gets to take a walk, one
gets to take a nap, and one gets to eat a cookie. And the
rat gets to choose who does each thing. So, one gets to
take a walk, right? Ask Ratty who he wants.
Child: Who do you want (*wanna) to take a walk?
Rat: I want the dog to take a walk.
The following video illustrates this performance.
Theories of language
acquisition
Imitation Theory: Children produce what they hear.
• Child: My teacher holded the baby rabbit and we patted
them.
Adult: Did you say that your teacher held the baby
rabbit?
Child: Yes
Adult: What did you say she did?
Child: She holded the baby rabbit and we patted them
Adult: Did you say she held them tightly?
Child: No, she holded them loosely.
Imitation Theory: Problems
• Non-imitation of parents
• Systematic errors across children and
languages: hitted, goed
(overgeneralization)
no drink (He does not want a drink)
dog toy (That’s the dog’s toy)
• Acquiring complex rules and producing
novel sentences
Reinforcement, explicit
instruction
Reinforcement Theory: Children learn
through explicit teaching with
positive/negative reinforcement
•
Child:
Want other one spoon, Daddy.
Daddy:
Child:
Daddy:
Child:
Daddy:
Child:
Daddy:
Child:
Daddy:
Child:
You mean, you want THE OTHER SPOON?
Yes, I want other one spoon, please, Daddy
Can you say ‘the other spoon’?
Other…one…spoon.
Say ‘other’
Other
‘Spoon’
Spoon
‘Other…Spoon’
Other…spoon. Now give me other one spoon?
Problems with explicit
instruction
• Parental reinforcement /instruction is seldom
seen
• Parents correct for meaning, not form (Brown &
Hanlon 1969)
•
Child: Mama isn’t boy, he a girl.
Adult: That’s right
Child: Walt Disney visits on Tuesday.
Adult: No, he doesn’t.
• Much of adult’s knowledge is NOT explicit
• When parents do correct, children don’t get it
Learning by analogy
Analogy Theory: Children learn language by
hearing a sentence and creating analogous
sentences
Problems:
• I painted a red barn.
I painted a barn red.
• I saw a red barn.
I saw a barn red (not possible, not produced)
• sun → sunny
moon → moony (not possible)
Innateness Hypothesis
• Humans are born ‘ready’ for language –
Universal Grammar is hardwired.
• Acquisition is rapid and completed by age fourfive
• Poverty of the stimulus – lack of negative
evidence and sometimes positive evidence
• Input may be fallible: speech errors, false starts,
ungrammatical and incomplete sentences
Exposure combined with UG leads children to
construct a grammar
Evidence for IH
•
•
•
•
universals in acquisition
children neither imitate nor are they instructed
innate perceptual categories
‘invention’ of language by children deprived of
real input
• critical age – normal cognition but abnormal
language
• abnormal cognition and normal language
(William’s syndrome, hydrocephalics)
• Inherited language deficiencies (Special
Language Impairment)
The Critical Period Hypothesis
The difference between adults and children
is in the ease of language learning – the
effect of ‘accent’, ‘grammatical mistakes’
etc. are attributed to the critical period
• Acquisition by the age of 7 yields native
command
• Acquisition between the ages of 8 and 15
yields progressively less perfect command
The innate ability to acquire language
degrades with age
Evidence for CPH: Feral
children
Genie, isolated until the age of 13.5, never learned
to produce more than telegraphic speech.
• Applesauce buy store.
• Neal come happy; Neal not come sad.
Isabelle, isolated until the age of 6.5, mastered
grammar within on year
• Why does the paste come out if one upsets the
jar?
• Do you go to Miss Mason's school at the
university?
CPH is not restricted to
language
Aspects of maturation are seen in humans and
animals.
• in ducklings: imprinting, ability to identify and
follow the mother
• in kittens: ability to perceive visual images
• in sparrows: ability to learn the father's songs
Developing the neural circuits for such skills is an
expensive allocation of developmental resources
and evolution favours individuals who lose this
once learning has (normally) occurred
Abnormal cognition, normal
language
Williams Syndrome (Chr. 11, IQ of 50):Cannot tie
shoe-laces, tell right from left, retrieve things
from the cupboard, add two numbers, find their
way, draw a bicycle. Understand complex
sentences and fix ungrammaticalities. Fondness
for unusual words.
Crystal (U. Bellugi): This is a story about chocolates. Once
upon a time in Chocolate World there used to be a
Chocolate Princess. She was such a yummy princess.
She was on her chocolate throne and then some
chocolate man came to see her. And the man bowed to
her and he said these words …
Inherited language deficiences 1
Specific Language Impairment (KE family, 3
generations, 30 members, 16 affected, M.
Gopnik)
Impairment uniform across affected members,
randomly distributed; Impervious to teaching and
correction; Persists through life; Problems with
inflections (plural, tense, agreement); non verbal
IQ tests normal, hearing normal.
• The boys eat four cookie
• Carol is cry in the church
• It’s a flying finches, they are
Inherited language deficiences 2
• No plausible environmental causes
• 53% of family but 3% of the total population
• Syndrome all or none, so likely one gene is the
cause rather than several genes, no graded
disability
• Autosomal gene – affects both sexes
• Dominant gene (not recessive) – one copy
Led to the isolation of the FOXP2 gene by Lai et al
which is implicated evolutionarily also in the leap
to language
Conclusion
Language experience is a mere "trigger"
Claim: children only use linguistic
experience to choose among a very
narrow range of possible hypotheses
given to them by the innate language
module – universal grammar.
Plato’s paradox is explained by the richness
of the innate structures.