Transcript Language Acquisition

Language Acquisition
5.
Elena Lieven,
MPI-EVA, Leipzig
School of Psychological Sciences, University of Manchester
LOT 5: 16-20 jan06
1
Grammatical Development
UG Predictions
U-B Predictions
across the board
emergence of
abstractions
within the domain of
a certain parameter
abstractions initially
local - then gradual,
piecemeal
abstractions based
on specifiable
characteristics of
input
LOT 5: 16-20 jan06
2
Factors affecting learning
•
•
•
•
Token frequency
Type frequency
Consistency
Complexity
• Prosodic information
• Semantic information
• Pragmatic information
BUT HOW ARE THESE INTEGRATED?
LOT 5: 16-20 jan06
3
Preferential looking vs Act-out or
Production
Children aged 1;9 and 2;1 look longer at a picture
that matches the correct agent of a novel action
that at another picture in which the same noun is
used for the patient but with a different novel
action
GIRL
NOVEL
ACTION 1
BOY
BOY
NOVEL
ACTION 2
GIRL
She’s blicking him
Gertner, Fisher & Eisengart, in press
LOT 5: 16-20 jan06
4
Does this conflict with the verb island hypothesis?
YES, in its strongest form
Does this reflect early and fully abstract knowledge of linking?
Not necessarily
LOT 5: 16-20 jan06
5
No verb is an island!
[McClure et al. in press]
• MLU of utterances with verbs in Stage 1
• MLU of utterances with old verbs in Stage 2
• MLU of utterances with new verbs in Stage 2
MLU old verbs stage 2 > MLU verbs stage 1
DEVELOPMENT OF KNOWN VERBS
MLU new verbs stage 2 < MLU old verbs stage 2
NEW VERBS HAVE LESS DEVELOPED ARGUMENT STRUCTURES
MLU new verbs stage 2 > MLU verbs stage 1
VALUE ADDED!!
NEW VERBS HAVE GAINED SOMETHING FROM THE KNOWLEDGE
OF OLD VERBS
LOT 5: 16-20 jan06
6
Auxiliary provision for BE and HAVE
[Theakston et al, 2005]
Mcr corpus: 2;0 – 3;0: provision still only a mean of 50% at 3;0
Specific subject-auxiliary combinations, e.g. : NP’s, he’s, it’s,
Proper-Noun’s, they’re, they’ve, I’m, I’ve, you’re, you’ve
Frequency in M’s speech predicts order of acquisition – except for you,
which is acquired late despite very high input frequency
Early acquired, high frequency, forms show higher levels of provision –
except for I, which is acquired early but shows very high omission
LOT 5: 16-20 jan06
7
It’s, He’s : - modelled frequently in input
- can be used in same way as adult
I/you : - have to be reversed
- child is more interested in talking about I than you
- maybe this overrides relative frequency of you
But why are there such high rates of omission with I ?
LOT 5: 16-20 jan06
8
Maybe children learn early on that I can substitute for you in
patterns with you learned from questions in the input
Mother: What’re you doing next?
Child: pink, orange, I doing orange
Mother: Are you drawing?
Child: I drawing
LOT 5: 16-20 jan06
9
Strengthening representations
[Chang, Dell & Bock, in press]
Discrimination task
Model could reliably (though not 100%)
discriminate after 4000 epochs
- can be done with partial representations
Production task
Model could produce transitive with novel verb
after 20,000 epochs
- needs a full representation
LOT 5: 16-20 jan06
10
Methodologies
LOT 5: 16-20 jan06
11
Experimental methods
• Allows us to pit one hypothesis against another
• Allows us to test children’s abstract knowledge
(with novel or low frequency items)
• One experiment never answers the question!
• Impossible to control everything in an
experiment
• What about the children who are dropped/don’t
respond?
LOT 5: 16-20 jan06
12
Novel transitives in 2-year-olds
Experiments give different ages for children’s ability to deal with novel verbs
in transitive constructions
• Fisher et al. Preferential looking
• Dittmar et al. Pointing and preferential looking
• Smith et al. Weird linking
• Chan et al., Preferential looking and Act-out
• Wittek & T Novel transitive production
These differences should be explicable by:
• partial representations being enough for some tasks but not others
• different tasks requiring different strengths of representation
• irrelevant task demands (often called performance limitations)
LOT 5: 16-20 jan06
13
Modelling
• Allows us to investigate distributional properties
• Allows us to play with the interaction between
different sources of information
• A lot depends on the precise way in which the
information is given to the model
• The model almost certainly doesn’t do it the
same way as the child
• So far not much modelling of semantics
LOT 5: 16-20 jan06
14
Naturalistic methods
• In principle, the participants are not
constrained
• Allows for detailed developmental
analyses
• Recording situations may be too limited
• One needs to use sophisticated analytic
tools to get beyond description
LOT 5: 16-20 jan06
15
Construction grammars
Advantages:
• Form-meaning mappings integral
• Cline from idiomatic through to abstract
• Explicit attempts to develop theory of
inheritance links between constructions
LOT 5: 16-20 jan06
16
Some issues
Are children learning constructions?
• Start with sub-parts [e.g.SV, VO, NPs]
• And perhaps with supra-parts [Discourse pairs]
Must children always learn form with function?
• Correct ungrammatical to grammatical but
semantically wrong:
– Kidd et al: Complement taking verbs
– Matthews et al: Conjoined NPs to correct WWO
– Abbot-Smith et al: weird VS intransitives to
semantically anomalous SVO
LOT 5: 16-20 jan06
17
Finally…….
• There is no doubt that frequency of various kinds plays a
role throughout learning and in the adult system
• There is no doubt that children’s language develops in
abstraction and complexity
• The relationship between characterising the adult system
(linguistics), explaining grammaticalisation and
explaining language development are in an increasingly
healthy relationship
• There is real hope of moving this scientific endeavour
forward in an exciting and productive way
Thank you!
LOT 5: 16-20 jan06
18