Producing Language

1
Producing Language

• Many partially overlapping processes (cascade)
• Planning different properties of different parts
  of message at any given moment
• For words coming up soon, planning sounds
• But for parts coming up later, still figuring out
  words (i.e., lemmas) sentence frame
• Havent gotten to sounds of far-ahead words yet
• How far ahead do you plan at the different
  stages?
• i.e., What are the sizes of the Planning Units?

2
Evidence from Speech Errorsabout Planning Units

• Properties of Word Exchanges
• The 2 words are usually similar in some ways
• Same syntactic category (both nouns or both
  adjectives or ...)
• From the same clause
• But dissimilar in other ways
• From different phrases
• They dont have to sound like each other to
  exchange
• Typically other words between them
• Properties of Sound Exchanges
• The 2 sounds are usually similar in some ways
• Same type of sound (both consonants or ... )
• From the same position in their word (both
  word-initial or ... )
• Typically from the same phrase
• But dissimilar in other ways
• Their words typically have different syntactic
  categories
• Typically no other words between the 2 words
  involved

3

• So, word exchange errors happen at a stage that
• Knows about syntactic categories of words
• Has ordered sentence frame with empty slots for
  words
• Each slot tagged for a particular syntactic
  category
• Frame is planned up to at least end of current
  clause
• It doesn't know about the sounds of the words
• Positional Level The slots are for lemmas
• Word exchanges happen when lemmas are put in the
  wrong slots in sentence frame
• But lemmas only fit into slots tagged with
  their syntactic category

4

• And sound exchanges happen at a stage that
• Knows about speech sounds
• Has ordered word frames with empty slots for
  sounds
• Each slot is tagged for a particular kind of
  sound
• e.g., initial consonant, vowel, coda consonant
• Frames are planned only up to end of current
  phrase
• It doesn't know about syntactic categories of
  words
• Sound Level The slots are for phonemes
• Sound exchanges happen when phonemes put in wrong
  phoneme slots in word frames
• But phonemes only fit into slots tagged with
  their phoneme-type

5
An Example with Multiple Errors

• The squeaky wheel gets the grease.
• gt
• The sqreaky guease gets the wheel.
• Notice, theres a sound exchange between 2 words
  that should have been far apart
• Sound exchanges are supposed to happen only
  between words that are closer together
• So, the word exchange must have happened first,
  in order for the 2 words involved in the sound
  exchange to be in position to be able to exchange
  their sounds
• Example provides evidence supporting the idea
  that words are ordered before their sounds are
  filled in

6
"Standard" Model of Language Production (Garrett)

• Series of cascaded stages
• Message Level Formulate a message to convey
• Functional Level Retrieve "words" (lemmas) to
  perform
• functions in message (agent,
• instrument, action, ...)
• Positional Level Build sentence frame that
  specifies
• where to put "words", given their
• functions
• Sound Level Retrieve sounds of words turn
  whole
• thing into a plan for articulation

7
Producing Language

• Cycle through series of stages over over
• Many partially overlapping processes (cascaded)
• Planning different properties of different parts
  of message at any given moment
• For words coming up soon, planning sounds
• But for parts coming up later, still figuring out
  words (lemmas) sentence frame
• Haven't gotten to sounds of far-ahead words yet

8
Sound Errors in Words

• Error outcomes are almost always legal for the
  language
• e.g., English doesnt have any words beginning
  with vl, English
• speakers never make slips like
• very flighty gt vlery fighty
• Furthermore, errors that result in saying real
  words are more likely than youd expect by chance
• barn door gt darn bore is more likely than
• dart board gt bart doard

9

• What does expect by chance mean here?
• For an error to result in saying wrong real
  words, there must be other words that are similar
  enough to the intended words
• i.e., to provide the opportunity for a word
  outcome
• e.g., barn door gt darn bore
• rotten cat gt cotton rat
• When you estimate how often such opportunities
  are likely to arise,
• given the vocabulary of the language
• Errors that result in words happen more often
  than they should, if they were due purely to
  chance
• Lexical Bias
• Its not that word outcomes are overall more
  likely than non-word outcomes

10
Top-Down Processing Again

• But maybe the lexical bias is on listeners
  side???
• Maybe we tend to hear errors as words if at all
  possible,
• even when the speaker actually produced a
  non-word
• Remember the phoneme-restoration effect?

11
A Technique for Inducing Sound Errors

• Present a series of word pairs
• ball doze
• bash door Interference Pairs Read silently
• bean deck
• bell dark
• darn bore Target Pair Say aloud fast
• Can't predict when you'll have to say a pair
  aloud, so prepare on all trials
• Possible responses
• darn bore No error
• barn door Exchange
• barn bore Anticipation
• darn door Perseveration
• Control the opportunities for word-producing
  errors
• Record the responses analyze them carefully
• Exchanges on about 30 of the critical trials

12
Some Results

• Exchanges resulting in word outcomes more
  likely
• ball doze big dutch
• bash door bang dark
• bean deck bill deal
• bell dog bark doll
• darn bore dart board
• barn door bart doard
• More likely Less likely
• Confirms perceived pattern in spontaneous errors
• Rules out Listener Bias as full explanation of
  Lexical Bias

13
Word Production Models

• All current theories of word production
• Explain why errors are usually similar in either
  sound or meaning to the intended target
• Have 2 stages
• 1. Retrieve lemma
• 2. Retrieve its sounds
• But they differ in
• How separate independent the 2 stages are
• Their mechanism for producing similarity effects
• Garrett's model vs Dell's model
• Modularity vs Interaction again!

14
Garretts Model of Word Production

• Lexicon organized into 2 files
• Meaning File
• Contains lemmas pointers to locations in Sound
  File
• Organized by meaning
• Sound File
• Contains word pronunciations
• Organized by sound

15

• To say a word in Garretts model
• Intended meaning
• Look in Meaning File and find lemma CAT
• Use CAT's pointer to find its pronunciation
  /kaet/ in Sound File
• Once you go into Sound File, youre done
  selecting which word to say (i.e., which lemma to
  choose)
• So what you find in Sound File cannot affect
  lemma choice

16

• In Garretts model
• Whole-word errors come from over- or under-shoot
  in Meaning File
• In right neighborhood, so find something similar
  in meaning
• Sound errors come from over- or under-shoot in
  Sound File
• In right neighborhood, so error should sound
  similar /kaeb/
• Garretts model was intentionally built with
  independent meaning sound stages
• Specifically to explain why errors seem to be
  related in one or the other way but not both

17
Mixed Errors

• Errors that are similar in both meaning and
  sound to intended word
• CAT gt rat
• ORCHESTRA gt sympathy
• In Garretts model, theres no way for both
  factors to interact in causing the error
• Something that looks like a Mixed Error is really
  just meaning-related error or just sound-related
  its a coincidence that its similar in the
  other way, too ( CAT gt rat )
• Or there were 2 independent errors, 1 at each
  stage
• ORCHESTRA gt SYMPHONY
• SYMPHONY gt sympathy
• Mixed Errors rare because coincidences double
  errors are rare

18

• Dell disagrees
• English vocabulary provides very few
  opportunities for Mixed Errors
• Pairs of words that are similar in both sound and
  meaning like cat rat or orchestra sympathy
  are very rare
• When you take that into account, Mixed Errors
• Happen more often than you would expect by chance
• Dells model was built to explain why errors tend
  to be related in
• Either sound or meaning or both

19
Localist
20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
(No Transcript)
26
(No Transcript)
27
(No Transcript)
28
(No Transcript)
29
Garrett vs Dell

• Meaning- or Sound-related errors
• Both models explain these
• Mixed errors
• Garrett's model explains why these are unlikely
• While Dell's model explains why they're
  especially likely
• They disagree about the data
• Legal outcome bias
• Requires an extra process in Garrett's model
• Pre-articulatory Editor (usually unconscious)
• Very likely to notice prevent illegal sound
  combinations
• Fairly likely to notice prevent non-words
• Less likely to notice an unintended word
• Natural consequence of architecture of Dell's
  model

30
Evidence for an Editor

• Motley, Camden, Baars (1982)
• shot home
• shame hear
• show hand
• hit shed
• People less likely to make errors resulting in
  taboo words
• Said unaware of possibility of saying taboo word
• But increased Galvanic Skin Response (GSR) on
  trials where there was an opportunity to say a
  taboo word

31
(No Transcript)
32
Testing Dells Model

• Lexical Bias caused by activation reverberating
  back forth
• Which takes time
• Prediction
• Errors should be less likely to be words as
  people talk faster
• Would be virtually impossible to observe with
  spontaneous errors
• But the prediction is confirmed when errors are
  elicited in the lab
• So, testing the models predictions led to the
  discovery of a new fact about speech errors
• Model is implemented as computer program (
  simulation) that talks
• Predictions derived from model are tested in
  studies with people