Onset Words and Phrases insure

Document Sample
Onset Words and Phrases insure Powered By Docstoc
					What do we need to do, to
understand an utterance?

                 Cartoon-head figures
                 from Jackendoff (1994),
                 Patterns in the Mind
       Speech Perception

Segment the auditory stream into
  words, made up of particular
 – What are the
 – Where are the word
              Word Recognition
Recognize individual words, resolving ambiguities

• Meaning
   Chris walked near the bank.

• Syntactic category (noun, verb, etc.)
   She saw her duck.
   Buffalo buffalo buffalo buffalo.

• Determine the structure
     of the sentence
   – Constituents
   – Hierarchical structure

• Buffalo buffalo buffalo buffalo.
• Put the ball in the box on the table.
            (ball in the box) on (the table)
            (ball) in (the box on the table)
     Semantic & Pragmatic
• What does the sentence mean?
  – Who did what to whom?
  – Truth conditions

• What is the speaker trying to convey?
  – Can you pass the salt?
    The Big Question:
  How Do We Accomplish
Linguistic Communication?

               Jackendoff (1994), Patterns in the Mind
   Map sounds to stored

                   1. A bird was in the
                      tree yesterday.
                   2. Are there any
                      birds in the tree?
                   3. A bird might be
                      in the tree.
                   4. That tree looks
                      like a bird.

What do we need?
Linguistic Knowledge as
  Categorical Rules for

    Jackendoff (1994), Patterns in the Mind
           Related Questions
• How do those categorical rules map onto cognitive
• Are the same cognitive processes used to produce
  speech as to understand it?
• Does modality (reading/listening) matter?
• Are our brains specialized for language?

 If I form a hypothesis about language
  understanding/production, how would I test it?
    What would count as data?
    Let’s give it a shot…
• Philosophical interest in language processing
  & language acquisition goes back to ancient
  – E.g., Aristotle on relations among thought,
    language, & external world
• Modern experimental approach quite new
  (most paradigms developed within last 50
• Modern theories reflect contributions from
  modern linguistic theory, cognitive
  psychology, computer science, & cognitive
  – Human language is unique among animal
   Early Psychological study of
• Wundt (1832-1920)
  – One of psychology’s
    founding fathers
  – Primarily used
    introspection for
    studying mental
    behavior, but one of
    first to use RT
  – Published on language
    in 1911
     Wundt’s (1911) hypotheses
         about Language

• Sentence (defined intuitively) is the
  primary unit of language
  – “Leave!” is one; “Days of the week” is not
  – “I filled the water with bottle.”
• Production converts a thought into a
  sequential string of sounds
• Comprehension is simply the reverse
  What’s wrong with
  Wundt’s Approach?

• Is introspection a reliable,
  replicable, objective tool for science?
• Do Wundt’s hypotheses lead to any clear
  predictions about behavior?
• Constructs (e.g., sentence) not carefully defined
• The stimuli/inputs for production and the
  response/outputs for comprehension are neither
  well-defined nor directly observable.
  – How can one develop a hypothesis and test it
   Would a different approach be
        more productive?

• Need to develop hypotheses that lead to
  clear predictions about behavior
  – Link environmental conditions to observable
• Need experimental techniques that can be
  clearly described and replicated in
  different laboratories.
Dominant paradigm in
psychology 1927-1960.
(Skinner, Pavlov)
  – How often does a behavior
    occur and with what
  – All behavior shaped by the
    environment using classical
    and operant conditioning.
  – No mental representations
  – Introspection devalued
           Skinner’s (1957)
           Verbal Behavior
• Language is a difficult phenomena
     for a behaviorist account.
  – In 1934, at a dinner party, philosopher A. N.
    Whitehead challenged Skinner to “account for my
    behavior as I sit here saying ‘No black scorpion is
    falling upon this table.’”
  – Skinner began the book the next morning, and spent
    20+ years working on it.
• Skinner often called this book his most
  important work
           Skinner’s (1957)
           Verbal Behavior

• Emphasis on production, rather than
• A sentence is a chain of associative links, “like
  beads on a string”
   There – is – no – black – scorpion…
• Speech is learned response to environmental
  stimuli (reinforcement, punishment)
    Experimental Evidence
• Speech is learned response to
  environmental stimuli (reinforcement,
  – Use of plural nouns increase if reinforced
    with “mmm-hmm” (Greenspoon, 1954,
  – Proportion of opinion statements increase
    if paraphrased/agreed (Verplank, 1955)
   Language research during the
      Reign of Behaviorism

• Operant studies (e.g., Verplanck)
• Classical Conditioning experiments (e.g.,
• Practical research, much of which was
  funded by the defense department
  – George Miller: understanding speech in noisy
    radio transmissions
         George Miller’s Lab
• Interested in speech and hearing
• Trained as a behaviorist in the 1940s
• In 1950’s investigated radio-based
  – How high does the signal-to-noise ratio need
    to be, for adequate transmission of the
     • Amount of noise
     • Characteristics of the message
     • Characteristics of the speaker
  How do we (the military) insure
adequate transmission of message?

 One strategy: Limit the
  vocabulary/possible messages
   – Digits are easy: 0-9 have 8 different
     nuclear vowels (only 5 and 9 share their
   – Nonsense syllables are opposite
     extreme—need to hear each phoneme
      George’s Ground-Breaking
Miller et al. (1951)

                       Miller & Selfridge (1950)
                       demonstrated analogous
                       pattern in free-recall test.

                       Why are words in sentences
                       easier to perceive and
                       easier to remember than
                       words in lists?
 What does “sentence-advantage”

Miller et al. (1951) maintain that sentences
  effectively restrict the number of alternative
  words, similarly to small vocabularies.

  “In 1951, I apparently still hoped to gain
  scientific respectability by swearing
  allegiance to behaviorism. Five years later,
  inspired by such colleagues as Noam
  Chomsky and Jerry Bruner [a social
  psychologist], I had stopped pretending to
  be a behaviorist.” (Miller, 2003)
 Miller (1962) re-examines Miller et
              al. (1951)
• Words in a sentence are not as distinct as words
  in isolation…
  – Less carefully pronounced (splice-test)
  – Words run together
• Why is there no extra cost for these?
• Speech rate of 2-3 words per second leaves little
  time for deducing set of alternatives after each
• “Reduction of alternatives” explanation is
Miller (1962)
Miller (1962)
Evolution of Speech and
      August 4, 2009
How did human vocal tract evolve?
All mammals produce vocal sounds in
   essentially the same way…
   Source – Oscillator (voicing) – Filter
   YouTube - vocal tract model synthesis
   YouTube - Vocal formants
In human speech, formants are the
   most informative parameter. They
   make speech intelligible.
    E.g., whispered speech lacks voicing and
     pitch, but has normal formants
Human speech requires fine, rapid
  motor control during articulation
   Role of formants in animal

Primates & birds perceive formants as
  accurately as humans
  – Individual identification via vocal
  – Provide cues to body size of “speaker”
    Diff’s btwn ape & human vocal tract

1. Human larynx lowers in throat during 1st yr of life
   ·Allows more tongue movement, for broad range of
     discriminable formant patterns
   ·Lowers formant freq-- impression of larger size
2. Human oral cavity
   shorter, nasal cavity
3. Humans lack
   laryngeal air sacs
   – Little known about
              Vocal Imitation
• Except for humans, primates are poor at
    • Apes raised like human kids
    • Monkeys raised with other species
    • Little evidence for learned vocal behavior

•   Humans clearly learn language(s)
•   Human whistling
•   Human bird calls
•   Human imitation of animal noises
            Vocal Imitation
• Whales, seals and dolphins are somewhat
  better than most primates.
  – Whales learn their songs

• Passerine Birds are terrific at this, even
  – Songbirds learn their songs
  – Mockingbirds learn other species songs, as well
    as environmental sounds (insects, car alarms,
  – Parrots can mimic human speech and even
    specific voices
  – Irene Pepperberg has trained African Grey Parrots
    to use human speech communicatively
Primates are poor candidates for
 production of spoken language
• Lack of rapid, fine motor control of
  vocal articulation
• Structure of the vocal tract
• Limited ability for vocal imitation
       Ape Language studies
              (1950’s to present)

• No luck training chimpanzees to produce
  spoken language
• Some success with manual/visual
• Chimpanzees, gorillas,
  & bonobos approximate
  linguistic skill of a 3-yr
  old human
• Ceiling on lg potential?
Koko (

Gorilla trained in sign language
 by Penny Patterson

Video Clip

 Has Koko acquired a
 language? What evidence is
 necessary to answer this
      Summing Up: How is human
         language special?
•   Vocal tract anatomy
•   Vocal imitation
•   Rapid, fine motor control of ariculators
•   Creative recombination of phonemes,
    morphemes, words for expression of nearly any

 But what about Koko?
 Compare Koko to Nicaraguan deaf kids
     Spontaneous emergence of
     Nicaraguan Sign Language
• Clip from “Birth of a Language”

• How is this signed communication the
  same/different from Koko’s?
• What kind of tests would you need to
  conduct to compare them?
Special Features of Human

• Specialized vocal tract: Broad range of
  formants for producing many distinct
• Vocal imitation/Social Learning
• Rapid, fine articulation
• Hierarchical structure
• Rule Learning
 Hierarchical Structure of Lg
Human speech has hierarchical structure,
 which is necessary to produce utterances of
 arbitrary complexity. Structure is distinct
 from content (specific phonemes).

  Syllable = (onset) + rhyme
  Rhyme = nucleus + (coda)
  Onset = one or more consonants
  Nucleus = one or more vowels
  Coda = one or more consonants
  Compositionality and the Rate of
       Data Transmission
• Small set of phonemes can be recombined very
  productively (but in a constrained way) to form
   – Morpheme = one or more syllables (meaning unit)
   – Signed language morphemes are also made up of “phonological”
     constituents (e.g., hand shape, movement, location)
• Morphemes can be combined productively (but
  constrained) to create words.
   – (prefix) + stem + (suffix)
   – Stem = (prefix) + stem + (suffix)
• Words can be combined productively to create
  Hierarchical Structure of Lg

Syntactic Hierarchies & Center-
• The man read Chaucer.
• The man who the woman despised read
• The man who the woman the children
  loved despised read Chaucer.
    What are the preconditions for
   learning hierarchical structure?
1. Fixed sequences (linear order):
   – Idioms & stock phrases (once upon a time) are fixed word
   – Words are fixed phoneme sequences
2. Statistical Learning is probably important for:
   – discovering words in speech stream
   – identifying syntactic category and subcategory of words
   – resolving lexical and syntactic ambiguity
   – Within phrases (e.g. NP), there is a predictable ordering of
     categories (e.g., the predicts a noun in the next word or two)

Predictive constraints on sequences may allow us to
  learn hierarchical relationships
 Coding the
  aids word
  n prior to

Jenny Saffron & colleagues
                             Marc Hauser & colleagues
Shared neural underpinnings to
 syntax & sequence learning?
• Broca’s aphasics who have severe syntactic deficits
  also exhibit deficits in sequence learning
  (Christiansen et al., 2001 unpubl)
• Incongruent musical sequences elicit P600’s, just
  like syntactic anomalies (Patel et al., 1998)
• MEG shows that Broca’s area is involved in
  processing music sequences (Maess et al., 2001)
• All higher organisms must learn about sequential
  events. How does human sequential learning
  compare with that of other primates?
              “The rat the cat the dog
                bit chased died.”

        What limits the kinds of
         rules that Tamarins can
         learn about sequences?
Fitch & Hauser (2004) suggest that
Tamarin’s can master Finite State
Grammars, but not Phrase Structure
         Types of Grammars
 All human languages allow for an infinite
 number of different utterances. What kinds
 of grammars allow this?

Finite State Grammars: A finite number of
  states (e.g., words, calls, syntactic
  categories), with rules for getting from one
  state to the next.
               class       sucks

                A+      really
FSG’s provide rules for concatenation
            Types of Grammars
A phrase structure grammar allows for long
  distance dependencies.
 (Last year, (Demi Moore (took (that cute dumb guy who’s about 20
   years old from “That 70’s Show”) out for a while))).

NP took out NP.
NP took NP out.
NP = NP + PP
NP = NP + S               The intervening NP can have an
                          arbitrary amount of internal
         Types of Grammars
Phrase structure grammars allow for center
 embedded constructions:

((The rat ((the cat (the dog bit _cat ))chased
  _rat)) died.)

The water someone I know carried spilled.
                  F&H 2004
A simple FSG could have two categories of
  “words”, A & B, with the rules that A must follow
  B and vice versa. ABn

A: no ba la wu, etc. (hi pitch)
B: li pa mo, etc. (low pitch)

No li ba pa                Tamarins & humans easily
Ba pa ba pa ba pa…         learn the simple finite state
*No ba la mo               grammar. In F&H, were the
                           learning conditions
                           comparable across the 2
                              F&H 2004
A simple phrase structure grammar might require
  equal numbers of A and B syllables. AnBn

A: no ba la wu, etc. (hi pitch)
B: li pa mo, etc. (low pitch)

((The rat ((the cat (the dog bit _cat ))chased _rat)) died.)
The water someone I know carried spilled.

What were the results for humans and tamarins in
 the PSG condition? What are the implications?
Why does human performance
surpass monkey performance?
1. Humans have UG
2. Humans have general cognitive abilities
   superior to monkeys (Look for evidence of
   this in non-language context.)
  –   Sequences can be learned via linear associations
      (as in FSG) or by learning ordinal positions of
      items (e.g., syllable structure, word position in a
      • Ordinal sequence learning in rhesus monkeys (Chen et
        al., 1997)
      • Strategies for nesting cups
• Monkeys learned 4 sets of        Chen et al. (1997)
  4 pictures each on touch-
• Trained to press pics w/in
  each set in a particular
  order. (Spatial config
• Then the 16 pics were
  reorganized into new sets
  & monkeys re-trained.
   – In “maintained” sets, the
     pictures were in the same
     ordinal slot [A,B,C,D].
   – In “changed” sets, the
     pictures were in a new
     ordinal slot.
   – Maintained sets were much
     easier to learn, suggesting
     that ordinal position had
     been encoded.
strategies: do
 they involve
   Human speech probably
   evolved as a result of …

• Rapid, fine motor control of articulators
  [frontal lobe, hypoglossal nerve]
• Ability to analyze sounds in terms of
  hierarchical structure [Broca’s area? UG?]
• Changes in the vocal tract & enhanced role of
• Increased ability to imitate auditory input
  [arcuate fasciculus?]

Shared By: