Transcript Slide 1

The cognitive psychology of
language – 1
• The previous lecture
covered mostly linguistics
– The study of languages as
independent structures
– Can apply to human languages as well
as machine languages
• Now focus on how cognitive
psychology can explain human
language
– Where is syntax stored? How are
resources used to apply the rules?
– How do we recognise words from
sounds?
1
Recognizing speech
• Recognition of words in speech
is extremely good
– Subjects recognized words after
270 msec of speech, even
though the words lasted 370
msec!
• Variations do not greatly affect
recognition
– Speed, volume and pitch have
almost no effect
– Recognition rate with completely
unfamiliar voice same as with
familiar voice
2
What affects recognition?
• More common words are
recognized faster
– ‘practice effect’
– Can pick a single phoneme from
common words much faster than
from uncommon ones
• Sentence context improves
recognition
– The same word in a sentence is
recognized faster that if it is
presented alone
– Expectation (top-down) helps
recognition process
3
Context &
Recognition
• Semantic priming contributes to
recognition
Teddy bears are good for little children
Teddy bears are good for hungry children
• ‘children’ will be recognized faster in
the first sentence
– The beginning of the sentence limits the
possible options
• Semantic priming can be quite a
strong effect (‘sentence superiority’)
As the sun set and evening fell, the pale
mcqn rose over the horizon.
4
Context is
cross-modal
• It does not matter if the priming is
done via reading or speech
– You can mix ‘em up and it still works.
– Shows it works at a language level
rather than a vision or auditory level
• Play the priming sentence via
headphones, show three words in
writing (Swinney, 1979)
– One unrelated word
– One word related semantically to the
entire sentence
– One word semantically related to the
last word of the priming sentence
5
Swinney Example
• Audio: “The aircraft was fueled, and
the crew had finished filing a flight
plan.”
– Words shown: PILOT, TOMATO,
BUILDING
– Ranking of speed recognition: PILOT,
BUILDING, TOMATO
• PILOT was recognized fastest due
to being primed by the sentence as
a whole
• BUILDING was only primed by the
last word (plan), so was recognized
more slowly
• TOMATO was recognized slowest,
because it was not primed at all
6
Lexical Access
• Where are the words stored?
– The lexicon is the word storage
• Lexicon holds all aspects of a word
– Spelling, meaning, pronunciation, role
in syntax
• The lexicon is huge
– Average American female office
worker, age 50
known well: 30,000
known vaguely: 8,000
used often: 16,000
used occasionally: 15,000
• Very rare to find someone pausing
to think of a word
7
– Lexical access is extremely fast
Lexical access models
• Two classes of model
– Direct access
– Serial search
• Direct access models
– Posit that the lexicon contains
some sorting information
– ‘bookmarks’ to word features
– Allows considering multiple
words at once for a search
• Serial search models
– No additional information needed
– Order of storage may be
important
– Can only consider one word at a
time
8
The Logogen model
(Morton, 1969)
• A threshold based direct
access model
• Each item in the lexicon is a
logogen
– A ‘feature counter’ which
accumulates evidence
– Features can be sensory,
contextual, anything
• Logogens build up evidence
until a threshold is reached
– Then it ‘fires’ and the word is
recognized
9
Logogen example
Let’s say the logogens fire at a threshold
of 5
BAT
BATTERY
/b/ /a/ /t/
Round
Involved in sport
Flying mammal
Has wings
Hairy
Lives in the Dark
/b/ /a/ /t/ /e/ /r/ /E/
Round
Electrical
Guns
Can shock
Has wires attached
Eventually goes dead
“The light wouldn’t go on,
so I checked the wires, but
then it turns out that the
battery was dead.”
10
Discussion: Logogen
model
• Morton later adapted threshold
of logogens
– Included the idea of ‘activation
decay’ (lowered threshold)
– Logogens encoding common
words have lower thresholds
• Precursor to connectionist
model
– Lacking multi-layer architecture
(esp. hidden layers)
– No competition or inhibition
– No clear idea of learning (in
logogen content)
11
The cohort model
(Marlsen-Wilson, 1973)
• Sequential, on-line model of
word recognition
• The first two phonemes used
as a filter to select ‘initial
cohort’ of likely words
– As more phonemes are heard,
candidates that don’t fit are
discarded, and the cohort
shrinks until you only get one
word left
• Very similar to the T9
predicitive text input on your
cell phone (!)
12
Cohort model example
• You hear /e//r/
• Initial list
– “aircraft”, “error”, “erode”, “air”,
“erroneous”
• Next you hear /o/
– That rules out “aircraft”, “air”,
“erode”
• Next you hear /n/
– That rules out “error”, so it must
have been ‘erroneous’
13
Discussion of the cohort
model
• Context does not affect the
cohort
– The cohorts are the same
regardless of priming
– Priming helps recognition
because it speeds up linking the
word to higher representations
• The decision process is not
passive
– Logogen model – activation
increases with no effort
– In the cohort model, words must
be actively rejected from the
cohort
14