Modeling and theory Insights from the Syntagmatic-Paradigmatic Learner Barend Beekhuizen Leiden University & University of Amsterdam 22 September 2015 Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 1 / 32
Modeling and theoryInsights from the Syntagmatic-Paradigmatic Learner
Barend Beekhuizen
Leiden University & University of Amsterdam
22 September 2015
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 1 / 32
1 Overview of SPLRepresentationsProcessingLearning
2 Main findingsModeling issues, theoretical puzzlesComprehensionProduction
3 Competence and performanceGaps in the theoryComprehensionRepresentationProduction
4 Wrap-up
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 2 / 32
Overview
1 Overview of SPLRepresentationsProcessingLearning
2 Main findingsModeling issues, theoretical puzzlesComprehensionProduction
3 Competence and performanceGaps in the theoryComprehensionRepresentationProduction
4 Wrap-up
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 3 / 32
Overview
The general problem
Mapping form to meaning
Acquisition:
Arriving at adult stateExplaining developmental waypoints
Starting point: Usage-Based framework
Computational cognitive model as method
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 4 / 32
Overview
The Syntagmatic-Paradigmatic Learner
Flow of the model:
1 Model receives input item: pair of an utterance and a number ofsituations
2 Model tries to analyze using processing mechanisms and existingrepresentations
3 Model updates grammar using learning mechanisms and best analysis
4 goto 1
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 5 / 32
Overview Representations
Representations
Constructions: pairings of signifiers and signifieds, both for ‘grammar’ and‘lexicon’
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 6 / 32
Overview Representations
F: ball
K
{object,entity,ball}
sd
srsr1
count = 4
F: Ɛ
K
F: go
K
{move}
{agent,mover} {location,goal}
{animate} {surface}
sd
srsr1 sr2
count = 12
Figure: Constructions
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 7 / 32
Overview Representations
F: ball
K
{object,entity,ball}
sd
srsr1
count = 4
F: Ɛ
K
F: go
K
{move}
{agent,mover} {location,goal}
{animate} {surface}
sd
srsr1 sr2
count = 12
Figure: Constructions
1 [ ball / ball ]
2 [ [ animate ] [ move / go ] ] |move(agent(animate),location(surface))
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 7 / 32
Overview Processing
Representations
Constructions: pairings of signifiers and signifieds, both for ‘grammar’ and‘lexicon’
Processing
An utterance in a situational context is analyzed using the set of knownconstructions and processing mechanisms
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 8 / 32
Overview Processing
{move}
{agent,mover} {location,goal}
{animate,Adam} {surface}
PHON: Adam
SEM:
PHON: ɛ
SEM:
{move}
{agent,mover} {location,goal}
{animate,Adam}
situation
O =
PHON: put
SEM:
PHON: it
SEM:
{move}
{patient,moved} {location,goal}
{object,entity} {surface}
{patient,moved}
{object,entity}
PHON: Adam
SEM:
PHON: ɛ
SEM:
{move}
{agent,mover} {location,goal}
{animate,Adam}
PHON: put
SEM:
PHON: it
SEM:
{move}
{patient,moved} {location,goal}
{object,entity} {surface}
Leftmost open constituent of this construction, pointing to {move}-node of meaning
Figure: Combine
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 9 / 32
Overview Processing
bootstrap
w3
ignore
meaning
c1
c2
meaning
c1
w1
w2
w4
c1
meaning
concatenate
Figure: Concatenate, bootstrap, ignore
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 10 / 32
Overview Processing
Representations
Constructions: pairings of signifiers and signifieds, both for ‘grammar’ and‘lexicon’
Processing
An utterance in a situational context is analyzed using the set of knownconstructions and processing mechanisms.Often many analyses possible, so find best one:
Most frequently encountered constructions
With fewest concatenate, bootstrap, and ignore operations.
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 11 / 32
Overview Learning
Representations
Constructions: pairings of signifiers and signifieds, both for ‘grammar’ and‘lexicon’
Processing
An utterance in a situational context is analyzed using the set of knownconstructions and processing mechanisms.Often many analyses possible, so find best one:
Most frequently encountered constructions
With fewest concatenate, bootstrap, and ignore operations.
Learning
Best analysis leaves trace in memory: 5 learning mechanisms.
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 12 / 32
Overview Learning
ball go there
a7
{event,move}
{role,patient,moved} {role,location,goal}
{entity,object,ball}
{entity,object,ball}
{event,move}
{role,location,goal}
{role,patient,moved}
K
F: ball F: go
K
F: ball
K
F: go
K
mccs
Figure: Adding most concrete constructions
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 13 / 32
Overview Learning
ball go there
a7
{event,move}
{role,patient,moved} {role,location,goal}
{entity}
{entity,object,ball}
{event,move}
{role,location,goal}
{role,patient,moved}
F: ε
K
F: go
K
F: ball
K
F: go
K
mcucsc
2
c4
c5
Figure: Updating the most concrete used constructions
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 14 / 32
Overview Learning
ball go there
a5
{entity,object,ball}
{event,move}
{role,location,goal}{role,patient,moved}
F: ball
K
F: go
K
csyn
Figure: Syntagmatization
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 15 / 32
Overview Learning
{entity,object,ball}
{event,move}
{role,location,goal}{role,patient,moved}
F: ball
K
F: go
K
{entity,object,ball}
{event,fall}
{role,patient,move}
F: ball
K
F: fall
K
{entity,object,ball}
{event}
{role,patient}
F: ball
K
F: ε
K
cpara
c c'
Figure: Paradigmatization
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 16 / 32
Overview Learning
{event,move}
{role,location,goal}
{entity,object,box}
st
{role,patient,moved}
{entity,object,ball}
ball go there
at
{event,state}
{role,location,goal}
{entity,object,floor}
st-1
{role,theme}
{entity,object,cup}
cup lies there
at-1
{role,location,goal}
F: there
K
cxsl
Figure: Cross-situational learning
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 17 / 32
Main findings
1 Overview of SPLRepresentationsProcessingLearning
2 Main findingsModeling issues, theoretical puzzlesComprehensionProduction
3 Competence and performanceGaps in the theoryComprehensionRepresentationProduction
4 Wrap-up
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 18 / 32
Main findings Modeling issues, theoretical puzzles
SPL resolves some a priori issues (chapter 2)
Comprehensiveness: comprehension and production
Simultaneity: lexical and grammatical constructions
Reappraisal of the starting-small approach
Learning as by-product of processing (immanence)
Reappraisal of the competence-performance distinction
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 19 / 32
Main findings Modeling issues, theoretical puzzles
SPL resolves some a priori issues (chapter 2)
Comprehensiveness: comprehension and production
Simultaneity: lexical and grammatical constructions
Reappraisal of the starting-small approach
Learning as by-product of processing (immanence)
Reappraisal of the competence-performance distinction
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 19 / 32
Main findings Comprehension
Comprehension (chapter 5, 6)
Robustness: making sense of utterance despite knowing little (usingconcatenation, bootstrapping)
Increasing coverage of utterance and situation
Increasing accuracy of picking out situation from 6 candidates
Varying mechanisms: XSL precedes bootstrapping; bootstrappingdominates.
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 20 / 32
Main findings Production
Production (chapter 7)
Experiment: give model situation, ask to produce utterance
Increasing length of produced utterance
Hardly any errors of comission
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 21 / 32
Competence and performance
1 Overview of SPLRepresentationsProcessingLearning
2 Main findingsModeling issues, theoretical puzzlesComprehensionProduction
3 Competence and performanceGaps in the theoryComprehensionRepresentationProduction
4 Wrap-up
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 22 / 32
Competence and performance Gaps in the theory
Linguistic knowledge grounded in language use
So we can reason from child’s productions to its knowledge oflanguage
However:
Sample may not contain reflection of full potentialOther reasons for not producing some linguistic itemInteractivity of components invalidates line of reasoning
So: need to account for a linguistic competence and performancewithin Usage-Based framework.
And show its explanatory value.
Not unique to SPL: all UB computational models do so. However,interaction lexical/grammatical acquisition gives interesting effects
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 23 / 32
Competence and performance Gaps in the theory
Linguistic knowledge grounded in language use
So we can reason from child’s productions to its knowledge oflanguage
However:
Sample may not contain reflection of full potentialOther reasons for not producing some linguistic itemInteractivity of components invalidates line of reasoning
So: need to account for a linguistic competence and performancewithin Usage-Based framework.
And show its explanatory value.
Not unique to SPL: all UB computational models do so. However,interaction lexical/grammatical acquisition gives interesting effects
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 23 / 32
Competence and performance Gaps in the theory
Linguistic knowledge grounded in language use
So we can reason from child’s productions to its knowledge oflanguage
However:
Sample may not contain reflection of full potentialOther reasons for not producing some linguistic itemInteractivity of components invalidates line of reasoning
So: need to account for a linguistic competence and performancewithin Usage-Based framework.
And show its explanatory value.
Not unique to SPL: all UB computational models do so. However,interaction lexical/grammatical acquisition gives interesting effects
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 23 / 32
Competence and performance Comprehension
Comprehension
Early abstraction
Increasing use of more concrete constructions
Does not entail loss of abstraction (to the contrary)
0.00
0.25
0.50
0.75
1.00
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
1000
0
time
prop
ortio
n n abstractslots
0
1
2
0.00
0.25
0.50
0.75
1.00
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
1000
0
time
prop
ortio
n n abstractslots
0
1
2
3
Figure: Abstraction of used length-2 and 3 constructions
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 24 / 32
Competence and performance Representation
Reflections of use on representations
Look under the hood to obtain a fuller understanding ofrepresentational potential
Abstractions are there, but not used so much
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 25 / 32
Competence and performance Representation
!"!"!!"#$#!"!%#&$#!#
"!%#&$&'()$#
"!%#&$$#
!"!"!*+,+'$#!"!%#&$#!#
!"!"!$%&'()$#!"!%#&$#!#
Figure: After 100 input items
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 26 / 32
Competence and performance Representation
!"!"!!"#$#!"!%#&$#!#
"!%#&$&'()$#
"!%#&$$#
!"!"!*+,+'$#!"!%#&$#!#
!"!"!$%&'()$#!"!%#&$#!#
!"!"!-$#!"!%#&$#!"!&'()$#!#
!"!"!!"#$#!"!&+.($#!"!/&$#!#
!"!"!!"#$#!"!%#&$#!"!/&$#!#
!"!"!$%&'()$#!"!%#&$#!"!%)*+*,$#!#
!"!"!!"#$#!"!-./'%00(1%$#!"!/&$#!#
!"!"!$%&'()$#!"!-./'%00(1%$#!"!%)*+*,$#!#
Figure: After 500 input items
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 27 / 32
Competence and performance Representation
!"!"!!"#$#!"!%#&$#!#
"!%#&$&'()$#
"!%#&$$#
!"!"!*+,+'$#!"!%#&$#!#
!"!"!$%&'()$#!"!%#&$#!#
!"!"!-$#!"!%#&$#!"!&'()$#!#
!"!"!!"#$#!"!&+.($#!"!/&$#!#
!"!"!!"#$#!"!%#&$#!"!/&$#!#
!"!"!$%&'()$#!"!%#&$#!"!%)*+*,$#!#
!"!"!!"#$#!"!-./'%00(1%$#!"!/&$#!#
!"!"!$%&'()$#!"!-./'%00(1%$#!"!%)*+*,$#!#
!"!"!$%&'()$#!"!%#&$#!"!(23%-*$#!"!4(-5&(4%!#!#
!"!"!$%&'()$#!"!%#&$#!"!(23%-*$#!"!/1!#!#!"!"!!"#$#!"!%#&$#!
"!&'()!$#!"!/1!#!#
!"!"!!"#$#!"!%#&$#!"!(23%-*!$#!"!/1!#!#
!"!"!*+,+'$#!"!%#&$#!"!&'()!$#!"!/1!#!#
!"!"!!"#$#!"!%#&$#!"!(23%-*!$#!"!4(-5&(4%!#!#
!"!"!!"#$#!"!%#&$#!"!/&!$#!"!/1!#!"!/&$#!#
!"!"!!"#$#!"!%#&$#!"!%)*+*,$#!"!/1!#!"!%)*+*,!#!#
!"!"!$%&'()$#!"!%#&$#!"!%)*+*,$#
!"!4(-5&(4%!#!"!%)*+*,!#!#
!"!"!$%&'()$#!"!%#&$#!"!(23%-*!$#!"!&'(,(!#!#
!"!"!$%&'()$#!"!-./'%50(1%$#!"!%)*+*,$#
!"!4(-5&(4%!#!"!%)*+*,!#!#
Figure: After 10000 input items
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 28 / 32
Competence and performance Production
Linguistic competence in production
Wysiwyg?
No:
Lexical items may be known but not produced because grammaticalconstructions are not known yetInteraction: competition between grammatical constructions
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 29 / 32
Competence and performance Production
The unexpressed expressables
Words that are known but nonetheless not produced, because there is (1)an erroneous word outcompeting them or (2) there is no grammaticalconstruction to ‘host’ them.
0
200
400
600
0 2500 5000 7500 10000time
coun
t
expressed
unexpressed unexpressable
unexpressed expressable
expressed
Figure: The expression of ‘second arguments’ over time
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 30 / 32
Wrap-up
Wrap-up
Why we need to focus on competence/performance
Corpora hide potential for abstraction
Simultaneity effects hide potential
Lexical knowledge hidden (unexpressed expressables)Paradoxal blocking effects
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 31 / 32
Wrap-up
Thank you
Barend Beekhuizen (Leiden & UvA) Modeling and theory 22 September 2015 32 / 32