The Competition Model Brian Mac Whinney- CMU Elizabeth

The Competition Model Brian Mac. Whinney- CMU Elizabeth Bates Janet Mc. Donald Kerry Kilborn Judit Osman-Sági Vera Kempe Yoshinori Sasaki Csaba Pléh Antonella Devescovi Takehiro Ito Jeffrey Sokolov Arturo Hernandez Michèle Kail Klaus-Michael Köpcke Ovid Tzeng Beverly Wulfeck Ping Li Empirical Results Published in: Mac. Whinney, B. , & Bates, E. (Eds. ) The crosslinguistic study of sentence processing. New York: Cambridge University Press, 1989. 15 articles since then 1

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 2

2. The Learner Distributed representations -> transfer Emergent modularity § § § Neuronal commitment, automaticity Capacity § § § Functional neural circuits Perspective-taking 3

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 4

1 A. Lexical Functionalism Form (cue, device) Function (role, meaning) 5

Competition between devices Competition between interpretations Agent Marking competition Patient Marking hidden Agent Function competition Patient Function 6

Cue validity -> cue strength Cues -> Interpretations Comprehension Meanings -> Devices Production pre agr init nom the giv def hidden act top per 7

Some cues The tiger pushes the bear. The bear the tiger pushes. Pushes the tiger the bear. The dogs the eraser pushes. The cat push the dogs. Il gatto spingono i cani. 8

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 9

Cue interactions • • • Peaceful coexistence Cue coalitions Competition between interpretations during comprehension Competition between devices during production Change from category leakage and reinterpretation 10

Cues vary across languages English: The pig loves the farmer SV > VO > Agreement German: Das Schwein liebt den Bauer. Den Bauer liebt das Schwein Case > Agreement > Animacy>Word Order Spanish: El cerdo quiere al campesino. Al campesino le quiere el cerdo. "Case" > Agreement > Clitic > Animacy > Word Order 11

Exotic Patterns Navajo: *Yas lééchaa’í yi-stin. snow dog him-frooze. Lééchaaa’ yas bi-stin dog snow him-frooze 7 -level hierachy of Animacy -- switch reference 12

Basic results § § § Reliable Cues Dominate Cue Strengths Summate Competition Cells show most variability 13

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 14

English Word Order 15

Italian Agreement 16

English Children 17

Hungarian Children 18

Italian Children 19

Cue validity (low levels) § Task frequency F(task T) / F(all tasks) § Simple availability (relative availability of a cue for a given task) F(times when cue A is present) The cat chases the dog. § Contrast availability F(cue A present ^ cue A contrasts) The cat chases the dogs. 20

Cue validity (high levels) § Simple reliability Reliable if always leads to right functional choice F(cue A present ^ cue A contrasts ^ cue A correct) / F (cue A present^cue A contrasts) § Conflict reliability In certain contexts, one cue will be more reliable F(cue A conflicts with other cue ^ cue A wins) / F(cue A conflicts with any cue) § SA -> CA -> SR -> CR transition 21

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 22

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 23

Pronouns - an online example Mac. Donald and Mac. Whinney (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 0 msec delay when there is a gender contrast only. 24

Pronouns - implicit causality Mc. Donald and Mac. Whinney (1994) Probes presented at 4 Delay Times: D 1 * 100 D 2 * pro D 3 * 200 D 4 * end * Gary amazed Ellen time after time, because he was so talented. N 1 Probes: V N 2 filler referent non-referent distractor verb , because PRO predicate. Frank Gary Ellen amazed Joel admires Susan because she is so fabulous. 25

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 N 2 right at the pronoun for E-S verbs! 4. Standard Competition Model cue summations and competitions, all right when they should occur. 26

2. The Learner § § Distributed representations -> transfer Emergent modularity § § Neuronal commitment, automaticity Capacity § § Functional neural circuits Perspective-taking The black dog is going to the market with his owner. 27

Parasitic Learning -- Kroll “turtle” Translation route “tortuga” 28

The Revised Hierarchical Model Kroll & Stewart, 1994 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. http: //psyling. psy. cmu. edu/traducir/ 31

Problems with Transfer § Lexical concepts “sibling” in Dutch = brother or sister § Broadness of application of translation equivalents glass in English, vidrio or vaso in Spanish car - “achterbak” or “kofferbak” tree -“stam” or “boomstronk” body - “romp” snout - “slurf” 32

More Problems with Transfer • Grammatical expression of certain aspects of experience The boy had fallen from the tree and his dog was hovering over him § Semantic boundaries differ across languages prepositions (Ijaz, 1986) Germans under-emphasize contact and overemphasize movement for “on” German “auf” means “up” 33

Emergent modularity § Growing modules § § § Farah and Mc. Clelland Jacobs, Jordan, Barto Kim et al. f. MRI study 34

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. . . 35

Dutch. L 1 English. L 2 36

Japanese. L 1 English. L 2 37

English. L 1 Dutch. L 2 38

Dutch. L 1 English. L 2 39

Aphasics - Word Order 40

Aphasics - Agreement 41

Case in Croatian Normals 42

Case in Croatian Aphasics 43

Word Order in Production 44

Some generalizations Children learn the most valid cues first. § Aphasics preserve the most valid cues. They also rigidify on the strongest devices § L 2 learners attempt transfer, but then learn cues. They gradually reach L 1 levels of cue strength. § Connectionism predicts transfer. § 45

3. The Context Providing negative evidence 46

Word learning - Merriman 47

Recovery in syntax 48

Complex cases 49

Mac. Donald et al. 50

Mac. Donald et al. 51

Open issues § § Neuronal Commitment Social Identification Resonance Setting up Input Contexts 52

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. 53