Language (and Decomposition). Linguistics provides… a highly articulated “computational” (generative) theory of the mental representations of language.

Language (and Decomposition)

Linguistics provides…

• a highly articulated “computational” (generative) theory of the mental representations of language that are used to “account for”– judgments of meaning and well-formedness– linguistics productivity: new sentences, new words– language acquisition: the ability of children to

project beyond their linguistic experience

Analysis of Invariance vs Analysis of Variance

• Linguist: in a sentence like, “John saw Mary,” the subject is the one who sees and the object the one seen.

• Psychologist: In a picture-matching task, position of the name in a sentence, initial vs. final, was significantly correlated with behavior– i.e, explicitly contrast “John saw Mary” and

“Mary saw John” and do an experiment

[Slides from Bob’s Analogy lecture]

Another ExampleJohn loves Mary

Loves

John Mary

Lover Beloved

Loves

Mary John

Lover Beloved

Mary loves John

Yet Another Example

John loves Mary and Bill hates Sally.

WithoutRole Bindings

John loves Mary and Bill hates Sally.

WithRole Bindings

John lovesMary and Bill

hates Sally.

John Loves Mary Bill Hates Sally

Loves

John Mary

Hates

Bill Sally

And

Relational structure has been lost!

Lover Beloved Hater Hated

Conjunct 1 Conjunct 2

More Problems Representing Structure

• But role bindings alone don’t solve the problem of multiple instances of a predicate.

• Example– John bought the apples yesterday and the pears last week.– John bought the pears yesterday and the apples last week.

John bought the apples yesterday and the pears last week (incorrect):

John bought the pears yesterday and the apples last week (incorrect):

John bought the apples yesterday and the pears last week (correct):

Representing Recursive Propositions in LTM

• A system with role bindings and instances of a predicate can represent recursively embedded propositions easily…

Recursive Embedding Example 1

Loves

Mary John

Lover Beloved

Sam knows Mary loves John

Knows

Knower Known

Sam

Recursive Embedding Example 2

Hug

John Mary

Hugger Hugged

Mary kissing John caused John to hug Mary.

Cause

Effect

Kiss

Mary John

Kisser Kissed

Cause

Structured Representations:A Linguistic Example

• What does “rebuild” mean?• (1) After the hurricane, John rebuilt his

house.– Did John build the house to begin with?– Does(1) presuppose that anyone built the

house before it was destroyed?– I'm going to pass on the help with the

donation to help rebuild the Old Man of the Mountain.

The Old Man of the Mountain

Representation of “rebuild”

• Involves a number of parts:– Building activity– A caused “creation” as the result of building– The presupposition: end state of building

existed before• The linguist assumes that the complex

structural representation of these parts would underlie any use or comprehension of the word

Tension: The Status of Linguistic Representations

• In linguistics, a “syntactic” representation mediates between sound (or letters, the hand movements of sign, etc.) and meaning (interface with other cognitive functions)

• The linguist operates under the assumption that language comprehension and production (as well as the generation of judgments of well-formedness) requires in every instance the creation of syntactic representations

• The tradition in cognitive psychology is to suppose that the linguistic representations are “abstract” and either not necessarily or rarely constructed when people speak and understand language on a daily basis -- people can use strategies, statistics, analogy, etc. to map directly between sound and meaning.

Example: Morphological Decomposition

• amiability• Linguistic representation:

nadj

root adj nami able ity

• “able” predicts category, meaning and potentiation of -ity

Pinker in “Words and Rules”3 Theories of “Irregular” Morphology

• full decomposition:“gave” = GIVE + PST => [ give [ ø ] ]• dual route (=Pinker’s mediation between

linguists and psychologists)“gave” = the past tense of “give”“walked” = WALK + PST => [ walk [ ed ] ]• single route:“gave” = the past tense of “give”“walked” = the past tense of “walk”

From the linguist’s perspective, only the full decomposition model makes any

sense

• “gave” behaves as a complex form with respect to syntactic and morphological distribution:– “the walkeding” = “the gav(e)ing”– He walked/He didn’t walk/*He didn’t walked– He gave/He didn’t give/*He didn’t gaved

“Evidence” for Dual Route View

• Lack of surface frequency effects for regular inflection:– RT to “walked” correlates with stem

frequency of “walk” rather than surface frequency of “walked”

– While RT to “taught” correlates with surface frequency of “taught” rather than stem frequency of “teach”

• Lack of priming for irregulars– “walked” primes “walk” but– “taught” doesn’t prime “teach”– (“darkly” and “darkness” prime “dark” but

“darkly” doesn’t prime “darkness”)

On surface frequency…

• Consider the informativeness of the past tense ending with respect to the stem:– For most regulars, -ed does not predict stem

(except for very high frequency regulars, and for those one gets surface frequency effects)

– For irregulars, form of past tense predicts past tense and vice versa

On priming…

• “stripes” primes “lion” (through “tiger”)

• Clearly, “taught” should semantically prime “teach”

• Lack of behavioral priming between “taught” and “teach” demands an explanation in terms of a theory of the task (e.g., Lexical Decision)

• False equation of “memorized” with “frequency” with “irregularity” with “whole vs. composed”

• “Irregularity” within language is governed by principles of structural locality

• Frequency always matters, both for the minimal constituents of language and for composed constituents, regardless of “regularity”

CNS 2008 Poster