Bridging the gap between L2 speech perception research and phonological theory

1
Bridging the gap between L2 speech perception
research and phonological theory
Paola Escudero Paul Boersma (March
2002) Presented by Paola Escudero March 1,
2002 Optimality Theory and Phonological Theory
Seminay Utrecht Institute of Linguistics OTS
2
Introduction

• Linguistics and L2 speech perception
• Bridging the gap
• OT and L2 mainly production
• L2 OT and perception
• Aims
• Assumptions
• Research hypothesis

3
Aims

• Incorporate speech perception and phonological
  theory in an OT model of phonemic categorization
• Account for a more challenging case of L2
  categorization, based on the integration of
  multiple acoustic cues
• Account for L2-specific patterns
• Further test the model with computer simulations
  that use the initial state and the algorithm that
  comes with the theory

4
Assumptions

• OT can handle the knowledge behind L1 and L2
  categorization and their development
• Full Transfer for the L2 initial state L1
  constraints and L1 rankings
• Perception modes L1 and L2 are handle by two
  separate systems
• Full Access for L2 learning access to the
  Gradual Learning Algorithm and the same
  strategies use in L1 (e.g. category formation,
  category split, etc)

5
Hypothesis

• Stochastic OT together with the GLA constitutes a
  successful model of L1 and L2 phonological
  development
• This model can account for the following
  observations
• Listeners optimise their perception in accord
  with the productions they encounter
• First-language (L1) listeners arrive at an
  optimal perception
• Second-language (L2) listeners manifest
  L2-specific optimisation strategies

6
Overview of the paper

• Observation 1 perception/production dependency
  and L1perceptual optimisation
• 1.1-1.2 Production perception differences in
  the same direction
• 1.3-1.4 The optimal listener vs. real listeners
• Hypothesis testing for L1 OT modelling
• 2.1 The knowledge behind L1 optimal
  categorization
• 2.2-2.3 The development of L1 optimal
  categorization in simulations
• 2.4 The simulated listeners vs real listeners
• Observation 2 L2 vowel categorization
• 3.1-3.2 L2 perception/production dependency L2
  optimization
• Hypothesis testing for L2 OT modelling
• 4.1-4.4 L2 initial state, development,
  simulations, and comparison with real L2
  listeners
• Discussion

7
Our case

• The vowels in the contrast have two main
  acoustic/auditory differences
• F1
• Duration
• We examine the preference patterns of these cues
  (in production and perception) in two English
  varieties Scottish and Southern British English
• These varieties are the target languages for the
  Spanish learners of English that will be
  presented in later sections

8
Production and perception differences in the same
direction
9
Optimal listener

• To optimise perception, the listener has
  tominimise the probability ofperceptual
  confusion

• The optimal perception strategy, therefore,
  islikelihood maximisation, i.e.choose the most
  likely produced category,given a certain F1
  duration

10
Average production environment
Scottish Southern
Optimal perception (max. likelihood)
82.5
87.1
11
L1 Modelling

• Constraint set
• Modelling the knowledge behind optimal
  perception
• Modelling the development of optimal perception

12
Native English constraint set

• 260 Hz should not be perceived as /I/
• 260 Hz should not be perceived as /i/
• 500 Hz should not be perceived as /I/
• 500 Hz should not be perceived as /i/
• 60 ms should not be perceived as /I/
• 60 ms should not be perceived as /i/
• 180 ms should not be perceived as /I/
• 180 ms should not be perceived as /i/

and so on, for all F1 and duration values
13
Scottish optimal perception in OT
74 ms, 349 Hz 349 Hz not /I/ 74 ms not /i/ 74 ms not /I/ 349 Hz not /i/
/I/ !
? /i/
 
 
 
 
 
 
 
Southern optimal perception in OT
74 ms, 349 Hz 349 Hz not /i/ 74 ms not /i/ 74 ms not /I/ 349 Hz not /I/
? /I/
/i/ !
 
 
 
 
 
 
 
14
How is the knowledge acquired?

• Whenever the listener makes a categorization
  error, she applies a Gradual Learning Algorithm

 
 
 
74 ms, 349 Hz 349 Hz not /i/ 74 ms not /i/ 74 ms not /I/ 349 Hz not /I/
? /I/ ? ?
? /i/ !? ?
 
 
 
15
L1 simulations

• Initial state
• Fed with the production distributions
• F1 and duration 21 steps, 84 constraints (21
  21 2 categories)
• 1000 data per virtual month
• Initial state scores 50 correct

16
Comparisons and preliminary conclusion

• Optimal vs. real listeners
• Optimal vs. simulated listeners
• Simulated vs real listeners
• L1 modelling conclusion

17
L2 production/perception dependency

• 7 beginners
• The more advanced bimodal distribution

Scottish

Southern

Spanish

L1

L1

L2

9
D E

0

1


D M

0

0

4

DS

1

4

1

SD

0

4

1

S M

4

6

2
S E

15

5

4

18
L2 Optimisation

• Also to minimize probability of confusion
• Full Access and grammar copying
• Initial state L1 categories and system
• Not good enough, thus further optimisation
• Full Access to the GLA and to language universal
  strategies category boundary, formation, split
• L1 and L2 two separate grammars

19
A possible assimilation pattern

• Native S.English has
• 13 arbitrary symbols

Spanish S.English has 2x5 vowels (7 symbols)
I i
i i?
E e
e e?
a a?
?, ? A, ?
o o?
? ?, o
u u?
U u
20
L2 Modelling

• L2 speech perception generalization and our model
• Full Transfer copy of constraints, rankings
• Native Spanish constraints
• Initial grammars constraint set rankings
• Further development boundary shift and length
  contrast
• L2 simulations
• Comparison with optimal and real L2 listeners

21
Discussion

• Category reuse and the initial state
• One or two perception systems?
• Fossilisation
• OT modelling
• The Algorithm

22
Conclusion

• Our formal model for L2 phonemic categorization
  successfully accounts for the attested optimal
  categorization in L1 acquisition as well as for
  the attested sub-optimal patterns in L2
  acquisition,thereby providing the linguistic
  mechanism that underlies the generalizations
  forwarded by several previous models of L2 speech
  perception