Topics in Cognition and Language: Theory, Data and Models

Download Report

Transcript Topics in Cognition and Language: Theory, Data and Models

Topics in
Cognition and Language:
Theory, Data and Models
*Perceptual scene analysis:
• extraction of meaning
• events, causality, intentionality, Theory of Mind
*Representation of meaning
Operations on meaning representations
• reasoning, induction, analogy
Sensorimotor Interaction
• joint attentlon, imitation
*Mapping Language to meaning
Introduction
Central issues of language acquisition
• Outline two approaches to explaining
language acquisition
• Highlight the property of “recursivity”
• Discussion: identify studies that can
help choose between competing
hypotheses
• Briefly discuss future events
• Prepare the next meeting
Language Acquisition
Conceptual
Representation
(Meaning)
Sentence
Learning the associations
<sentence, meaning>
• Two problems
• Extracting meaning from the world (by
vision, audition, touch etc.)
• Learning the sentence-meaning relations
Scene
(Some) Data to account for:
Most of children's early language is
grammatical from the adult point of view
- simplified
- follows adult grammatical conventions
fairly well
• Children can generalize to use verbs in
ways they have not heard before
- produce some creative yet canonical
utterances that they could not have
heard from adults
The "Poverty of the stimulus"
Obstacle
The possible mappings of sentence
to meaning are uncountable
• Language acquisition is an
indeterminate problem
• The training data is underspecified
How to approach this
question?
What is the “initial state” or architecture
of the learning system?
• How does it learn?
• Data to account for
• Specify requirements for an artificial
system that can learn language*
Two Explanations
Nativist: Chomsky, Pinker:
• Children do not learn abstract syntactic
structures at all, but rather
• they already possess them as a part of
their innate language faculty (UG).
• Usage Based Learning: Tomasello, Goldberg:
• Early utterances organized around
particular concrete words and (idiom-like)
phrases, not system-wide syntactic schemas.
• Abstract and adult-like syntactic
categories and schemas emerge only gradually
and in piecemeal fashion during the preschool
years.
How they account for the Data:
Most of children's early language is grammatical
from the adult point of view
• Nativist: because the structure is already there
• Usage-based: Because they are performing
“idiomlike” copy and paste.
• Children can generalize to use verbs in ways they
have not heard before
• Nativist: because they exploit the innate adult
syntax
• Usage-Based: This generalization does occurs
later in development, suggesting learning.
Types of Generalization:
Lexical categories (e.g. Nominal substitution):
Take ___. Eat ___. Draw ___ on ___.
• Re-use of Grammatical constructions:
new verb learned in intransitive form
generalized to transitive form
• Creative constructions
recursivity
Predictions about transfer of grammatical forms:
• Nativist - once a gramatical form (e.g.
transitive) is present it should be fully available
• Usage-based - transitive is element-soecific,
theu becomes generalized
Transfer is progressive
Productive transitive
utterances in different
studies. Percentage of
children (or responses in
some cases – see Table 1)
that produced transitive
utterances of a novel verb
that was heard in some
other sentence frame. The
data points correspond to
the studies listed in Table 1
(Tomasello 2000)
Usage-Based Learning
1. Item based imitation
2. Generalization
• Across types (e.g.concrete nouns)
• Across structures (e.g. Transitive)
3. Creative Re-combination *
* What principles govern the ways in which children
combine established linguistic constructions with one
another creatively? Recursivity
Source of Recursivity
in Language
Chomsky:
• Syntax is source of recursive structure in
semantics
Jackendoff:
• Syntax and semantics are parrailel
recursive systems
• Meaning has complex combinatorial
structure that is not derived from syntax
Extreme alternative:
• Semantics is the source of recursive
structure
Parallel Recursive Structure
Major Issues
How does the system
• Extract meaning from the world
(by vision, audition, touch etc.)
• Learn the sentence-meaning
relations
1. Item based imitation
2. Generalization
• Across types (e.g.concrete nouns)
• Across structures (e.g. Transitive)
3. Creative Re-combination *
Language
Acquisition
Test-Bed
Upcoming Meetings
1. June: "Vision and Language: A Robotlcs
Perspective"
2. September: Mehler + Tomasello: Debate
3. Next Meeting:
• Data to accounz for?
• Recursion in language and meaning
- Hauser et al. 2002 Science
- Goldberg 2003 TICS
References
• Ray Jackendoff. Parallel constraint-based generative
theories of language, Trends in Cognitive Sciences, Volume
3, Issue 10, 1 October 1999, Pages 393-400
• Michael Tomasello. The item-based nature of children's
early syntactic development, Trends in Cognitive Sciences,
Volume 4, Issue 4, 1 April 2000, Pages 156-163