The Competition Model Brian MacWhinney- CMU

1
The Competition Model Brian MacWhinney- CMU

• Elizabeth Bates Csaba Pléh Michèle Kail
• Janet McDonald Antonella Devescovi Klaus-Michael
  Köpcke
• Kerry Kilborn Takehiro Ito Ovid Tzeng
• Judit Osman-Sági Jeffrey Sokolov Beverly
  Wulfeck
• Vera Kempe Arturo Hernandez Ping Li
• Yoshinori Sasaki
• Empirical Results Published in
• MacWhinney, B., Bates, E. (Eds.) The
  crosslinguistic study of sentence processing.
  New York Cambridge University Press, 1989.
• 15 articles since then

2
1. The Input

• A. Lexical Functionalism -- constructions
• B. Input-driven Learning -- cues, frequencies
• Cue validity predicts cue strength
• p(function)form - comprehension
• p(form)function - production

3
2. The Learner

• Distributed representations -gt transfer
• Emergent modularity
• Neuronal commitment, automaticity
• Capacity
• Functional neural circuits
• Perspective-taking

4
3. The Context

• Classroom context
• Negative feedback is positive feedback
• Instructional format interacts with learner
  characteristics
• Role of computerized instruction
• Setting up input contexts
• Role of lexical richness
• Learner must learn how to learn

5
1A. Lexical Functionalism

Form (cue, device)
Function (role, meaning)
6
Competition between devicesCompetition between
interpretations
Agent Marking
Patient Marking
competition
hidden
Patient Function
Agent Function
competition
7
Cue validity -gt cue strengthCues -gt
Interpretations ComprehensionMeanings -gt
Devices Production
pre
the
nom
agr
init
hidden
per
top
act
def
giv
8
Some cues

• The tiger pushes the bear.
• The bear the tiger pushes.
• Pushes the tiger the bear.
• The dogs the eraser push.
• The dogs the eraser pushes.
• The cat push the dogs.
• Il gatto spingono i cani.

9
The dog was chased by the cat.

• Comprehension - Interpretations compete
• Agent The dog vs. the cat
• Patient The dog vs. the cat
• Production - Devices compete
• Dog placement preverbal, postverbal, by-clause
• Cat placement preverbal, postverbal, by-clause

10
Cue interactions

• Peaceful coexistence
• Cue coalitions
• Competition between interpretations during
  comprehension
• Competition between devices during production
• Change from category leakage and reinterpretation

11
Cues vary across languages

• English The pig loves the farmer
• SV gt VO gt Agreement
• German Das Schwein liebt den Bauer.
• Den Bauer liebt das Schwein
• Case gt Agreement gt AnimacygtWord Order
• Spanish El cerdo quiere al campesino.
• Al campesino le quiere el cerdo.
• "Case" gt Agreement gt Clitic gt Animacy gt Word Order

12
Exotic Patterns

• Navajo
• Yas lééchaaí yi-stin.
• snow dog him-frooze.
• Lééchaaa yas bi-stin
• dog snow him-frooze
• 7-level hierachy of Animacy -- switch reference

13
Basic results

• Reliable Cues Dominate
• Cue Strengths Summate
• Competition Cells show most variability

14
Ungrammaticality

• Continuity for pockets of grammaticality
• Hungarian possessive for accusative
• Croatian neutralized case in masculine
• Japanese wa marking
• Slowdown for grammatical sentences in Russian,
  Hungarian, Spanish without the preferred cue
• Cue summation for pronominal processing

15
English Word Order
16
Italian Agreement
17
English Children
18
Hungarian Children
19
Italian Children
20
Cue validity (availability)

• Task frequency
• F(task T) / F(all tasks)
• Simple availability
• F(cue A present) / F (all cases of task)
• Contrast availability
• F(cue A present cue A contrasts)

21
Cue validity (reliability)

• Simple reliability
• F(cue A present cue A correct) /
• F (cue A present)
• Contrast reliability
• F(cue A present cue A contrasts cue A
  correct) /
• F (cue A presentcue A contrasts)
• Conflict reliability
• F(cue A conflicts with other cue cue A wins) /
• F(cue A conflicts with any cue)
• SA -gt CA -gt SR -gt CR -gt Conflict transition

22
Cue validity vs. cue strength

• Cue validity is based on (tedious) counts of
  texts
• Cue strength is first assessed through ANOVA
  analyses in Competition Model experiments
• Cue strength is then modeled using MLE

23
MLE models of cue strength

• P (first noun) ? S i (first) /? ? S j (others)
• Two choice case
• P (first noun)
• ? S i (first) /? S i (first) ? S j (second)
• Models vary number of parameters and can be
  additive or multiplicative

24
Pronouns - an online example

• MacDonald and MacWhinney (1989)
• Just before dawn, Lisa was fishing with Ron in
  the boat,
• and she caught a big trout right away.
• and lots of big trout were biting.
• Priming of referent at 500 msec for unambiguous
  gender.
• Slowdown in processing of probes right at 0msec
  delay when there is a gender contrast only.

25
Pronouns - implicit causality

• McDonald and MacWhinney (1994)
• Probes presented at 4 Delay Times
  D1 D2
  D3 D4
• 100
  pro 200 end
  Gary amazed Ellen time after time, because he
  was so talented.N1 V N2
  filler , because PRO predicate.Probes
  referent
  Gary non-referent
  Ellen distractor
  Frank verb amazed

26
Results and Competition

• 1. Slowdown in processing of probes at pronoun
  when there is a contrast.
• 2. Facilitation from pronoun onwards when first
  noun advantage agrees with implicit causality.
• 3. Activation of N2 right at the pronoun for E-S
  verbs!
• 4. Standard Competition Model cue summations and
  competitions, all right when they should occur.

27
2. The Learner

• Distributed representations -gt transfer
• Emergent modularity
• Neuronal commitment, automaticity
• Capacity
• Functional neural circuits
• Perspective-taking

28
Parasitic Learning -- Kroll

Translation route
turtle
tortuga
29
Transfer

• Principle Everything that can transfer will.
• Connectionism predicts transfer
• Word order can transfer
• Phonology can transfer
• Meaning can transfer
• Morphological markings cannot
• Early bilinguals as mixed

30
Transfer beyond the word

• I want to go to school.
• Yo querer ir a escuela.
• I would like to go to school.
• (I) would-like to-go to the-school.
• xx quer-rí-a ir a la-escuela.
• Do you want to eat at my house?
• You want not want at me eat, huh?
• Translation with feedback may not be so bad.

31
Emergent modularity

• Growing modules
• Farah and McClelland
• Jacobs, Jordan, Barto
• Kim et al. fMRI study

32
Capacity restrictions

• Detectability
• Complexity (for production)
• Assignability (memory load)
• Online load minimization
• One good cue is enough (Russian, Spanish)
• Waiting for a reliable cue Russian, Hungarian
• No use waiting for cue that will not be reliable,
• German die Frau küßt der ...

33
DutchL1 EnglishL2
34
JapaneseL1 EnglishL2
35
EnglishL1 DutchL2
36
DutchL1 EnglishL2
37
Aphasics - Word Order
38
Aphasics - Agreement
39
Case in Croatian Normals
40
Case in Croatian Aphasics
41
Word Order in Production
42
Some generalizations

• Children learn the most valid cues first.
• Aphasics preserve the most valid cues.
• They also rigidify on the strongest devices
• L2 learners attempt transfer, but then learn
  cues. They gradually reach L1 levels of cue
  strength.
• Connectionism predicts transfer.

43
3. The Context

• Providing negative evidence

44
Word learning - Merriman
45
Recovery in syntax
46
Complex cases
47
MacDonald et al.
48
MacDonald et al.
49
Open issues

• Neuronal Commitment
• Social Identification
• Resonance
• Setting up Input Contexts

50
Conclusions

• Models of Input, Learner, and Context must
  interlock
• Competition Model is properly accounts for what
  we know about language learning, but
• The model must be developed still further.