Visual Word Lecture

Download Report

Transcript Visual Word Lecture

Mental Lexicon
• All of your knowledge about words
– and you know a lot of words!
• Average college-educated adult
– Speaking vocabulary = 75,000 - 100,000 words
– Recognition vocabulary is substantially larger
• You're not equally likely to use all of those words
– The 50 most common words make up
• ~60% of the words we speak
• ~45% of the words we write
– On average, you only say 10-15 words before repeating one
9/15/10
Psyc / Ling / Comm 525 Fall 10
10 Most Frequent English Words
(counts out of ~1,000,000 words)
Written
the
be
of
and
a
in
he (she)
to (infin.)
have
to (prep)
Spoken
70,000
40,000
20,000 (6000)
11,000
I
and
the
to (?)
that
you
it
of
a
know
65,000
38,000
15,000
Notice that most are function words rather than content words
9/15/10
Psyc / Ling / Comm 525 Fall 10
Bottom-up & Top-down Processing in
Visual Word Recognition
9/15/10
Psyc / Ling / Comm 525 Fall 10
Writing Systems
•
Two basic types of writing systems
– Ones that indicate pronunciation
– Ones that do so less
•
Systems that do represent pronunciation:
– Alphabets - One character supposed to represent one “sound”
• ALL modern alphabets are derived from Phoenician
– Syllabaries - One character represents a whole syllable
• In most, 2 syllables that share a sound don’t look anything alike
– Japanese hiragana: か= /ka/ vs き= /ki/ vs く= /ku/
– But, Korean hangul has C & V characters within syllable: 수잔 = /suzn/?
•
Systems that represent pronunciation less directly:
– Ideograms (pictograms) - One character represents a meaning
• Words with similar meanings usually share characters, even if pronunciation
is completely different
• But often words that share a syllable that has different meanings in each
word also share a character
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
Word Boundaries
• Most (all?) languages using Roman alphabet put spaces between words
• But some other writing systems do not (e.g. Chinese, Japanese)
- Sometimes ambiguous where word boundaries are (just as in speech)
9/15/10
Psyc / Ling / Comm 525 Fall 10
Bottom-up and Top-down Processing
in Reading
• Some examples:
– Detecting particular letters is less accurate in highly familiar words
– Proofreading is harder the more familiar the text
9/15/10
Psyc / Ling / Comm 525 Fall 10
Does this remind you of
anything about auditory
word recognition ?
9/15/10
Psyc / Ling / Comm 525 Fall 10
Bottom-up and Top-down Processing
in Reading
• Some examples:
– Detecting particular letters is less accurate in highly familiar words
– Proofreading is harder the more familiar the text
• - ...
• Race Models of word recognition
– Top-down and bottom-up processes go on in parallel
– Racing with each other
– Whichever "finishes" first wins the race
• i.e. determines how you identify the word
– If bottom-up processing is hard because input is noisy, top-down wins
– If little help from context, bottom-up wins
• Decision Criterion = Finish line in race
– How sure must you be that the input is a word before saying so?
9/15/10
Psyc / Ling / Comm 525 Fall 10
Disorders of Reading
• Patterns of Acquired Dyslexia have influenced theories
and models of normal reading
– More than observations of any other kind of language deficit
have influenced models of other aspects of normal language
processing
9/15/10
Psyc / Ling / Comm 525 Fall 10
Surface Dyslexia
(tends to occur in fluent aphasics with posterior brain damage)
• Read regularly spelled words aloud ok
• Read nonwords aloud ok
• Tend to mispronounce irregularly spelled words
– They regularize them
• island > /Izlǽnd/
• pint > /pInt/
– So, they seem to
• construct pronunciations via direct letter-to-sound mappings
• without retrieving knowledge about particular words' pronunciations
9/15/10
Psyc / Ling / Comm 525 Fall 10
Phonological Dyslexia
(often no other aphasia)
(most similar of the acquired dyslexias to developmental dyslexias)
• Read highly familiar words aloud ok, regardless of spelling regularity
• Trouble reading both less familiar words and non-words aloud
– Tend to pronounce them as similar-looking familiar words
•
•
•
•
i.e., they lexicalize them
forb > fork
moth > mother
border > bread
– If the word they come up with happens to have an irregular
spelling for its pronunciation, they pronounce it in the correct
irregular way
– So, they seem to
• get into the neighborhood of words that look like what they see
• & retrieve the pronunciation of one of the more familiar words in that
neighborhood
9/15/10
Psyc / Ling / Comm 525 Fall 10
Deep Dyslexia
(tends to occur in non-fluent aphasics with anterior brain damage)
• Read content words aloud a lot better than function words
• Within content words, better on concrete, imageable ones
• Often can't read non-words at all, or may lexicalize them
• Errors sometimes semantically related, with no sound or spelling
similarity
– ape > monkey
– forest > trees
• Errors sometimes visually related instead, or mixed visual and
semantic
– scandal > sandals
– orchestra > sympathy
• So, they seem to (sometimes)
– get into the neighborhood of words with meanings like what they see
– & then retrieve the pronunciation of another word in that neighborhood
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
Modular vs Interactive Processing Systems
• It’s obvious that both bottom-up and top-down
processes contribute to the recognition of letters &
sounds & words
• But how does top-down processing work?
– Interactive account:
• Context & knowledge guide actual perception of input
vs
– Modular account (= post-perceptual, autonomous):
• Context & knowledge influence choices among alternative
candidates proposed by perceptual processes
9/15/10
Psyc / Ling / Comm 525 Fall 10
Localist
9/15/10
Psyc / Ling / Comm 525 Fall 10
Localist
9/15/10
Psyc / Ling / Comm 525 Fall 10
Localist
9/15/10
Psyc / Ling / Comm 525 Fall 10
Modular Account of Phoneme Restoration
• The connectionist account is interactive
• In contrast, a modular account says:
– No top-down feedback from words to sounds
– Instead, system guesses there must have been s because that's
what would make sense
– An unconscious decision process
9/15/10
Psyc / Ling / Comm 525 Fall 10
• Each of the next 3 slides has a list of letter
strings
• Your task is to read through them as quickly as
you can and count how many of them are words
• Raise your hand as soon as you’re done
9/15/10
Psyc / Ling / Comm 525 Fall 10
zyndc
cnccl
apple
frgtd
wrpts
brat
nxprd
must
lbdry
other
nrgln
sfbdl
war
cloth
dtrnp
library
stwsn
mplfs
9/15/10
Psyc / Ling / Comm 525 Fall 10
bant
anger
fold
bagin
pretser
mash
kalt
magic
lomp
sinos
arid
hink
radle
track
rean
supper
weth
amol
9/15/10
Psyc / Ling / Comm 525 Fall 10
brane
leev
want
damp
stane
mair
quick
lowd
heeter
power
wim
pryse
muther
prefer
koller
heaven
much
prufe
9/15/10
Psyc / Ling / Comm 525 Fall 10
• Why were you slower on the second list than on
the first, and even slower on the third list, even
though there were 7 words in each list?
– Because the nonwords (NWs) grew progressively
more word-like across the lists
• In list 1, the NWs were not even pronounceable
and had illegal sequences of letters
• In list 2, the NWs were pronounceable and
followed legal English spelling patterns
• In list 3, NWs all had the same pronunciation as a
real word
= pseudohomophones
9/15/10
Psyc / Ling / Comm 525 Fall 10
Tasks & Strategies
• You adopted different decision criteria in the 3
different lists
– about how fully to process the letter strings before moving on to
the next one
• So, sometimes the "distractors/fillers" in an experiment
can matter a lot!
– Influence task-specific response strategies people can adopt
– Crucial to think very carefully about how participants could be
doing the tasks we give them
– But also important to realize people’s intuitions about how
they’re doing something are often not reliable
9/15/10
Psyc / Ling / Comm 525 Fall 10
• The rest of the slides here are ones I didn’t get
to in class
9/15/10
Psyc / Ling / Comm 525 Fall 10
• When you encounter a new word you don’t know,
you can tell a lot about it from:
– its position in the sentence relative to other words you
do know
= Syntax
– its prefixes and suffixes
= Morphology
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
The Jabberwocky
Lewis Carroll
'Twas brillig, and the slithy toves
Did gyre and gimble in the wabe:
All mimsy were the borogoves,
And the mome raths outgrabe.
One two! One two! And through and through
The vorpal blade went snicker-snack!
He left it dead, and with its head
He went galumphing back.
"Beware the Jabberwock, my son!
The jaws that bite, the claws that catch!
Beware the Jubjub bird, and shun
The frumious Bandersnatch!“
"And hast thou slain the Jabberwock?
Come to my arms, my beamish boy!
Oh frabjous day! Callooh! Callay!"
He chortled in his joy.
He took his vorpal sword in hand:
Long time the manxome foe he sought -So rested he by the Tumtum tree,
And stood a while in thought.
'Twas brillig, and the slithy toves
Did gyre and gimble in the wabe:
All mimsy were the borogoves,
And the mome raths outgrabe.
And, as in uffish thought he stood,
The Jabberwock, with eyes of flame,
Came whiffling through the tulgey wood,
And burbled as it came!
Adjectives, Nouns, Verbs
= Parts of speech
= Syntactic categories
9/15/10
Psyc / Ling / Comm 525 Fall 10
Morphology
•
Many words have internal structure
•
Morpheme = smallest meaningful unit in language
•
Affixes = prefixes, suffixes, & infixes
•
Free morphemes (friend, the) vs
Bound morphemes (un-, -ly, -est)
•
Lexical (= Content) morphemes (friend) vs
Grammatical (= Function) morphemes (the, un-, -ly, -est)
•
Allomorphs = different versions of same morpheme
9/15/10
– friend > friendly > unfriendly > unfriendliest
– friend
– unfriendliest
= 1 morpheme
= 4 morphemes (end is not a morpheme in friend)
– (Remember allophones?)
– English plural = /s/, /z/, /Iz/, /In/, ...
– English indefinite article = a, an
Psyc / Ling / Comm 525 Fall 10
• Languages vary from having:
– Many short simple words and using more of them
• (e.g. Chinese = isolating)
– To having mostly long complex words and using few of them
• (e.g. Turkish, Hungarian = agglutinative)
• English is somewhere in between
• Term Morphosyntax reflects the fact that the same kinds
of relationships are coded morphologically in some
languages & syntactically in others
9/15/10
Psyc / Ling / Comm 525 Fall 10
Does Morphological Structure
Affect Reading?
• People do seem to decompose complex words during reading
– Priming from regularly inflected morphological relatives can be
equivalent to repetition priming
• e.g., believes primes believe just as much as believe primes itself
– But priming from irregularly inflected or derivational relatives is
smaller
• e.g., believer doesn’t prime believe as much as believe primes
itself, nor does went prime go as much as go primes itself or as
much as repeated primes repeat
• Even when what looks like a morpheme really isn’t
– e.g., beak- and -er in beaker
– Get effects of frequencies of apparent subcomponents
• The frequency of the word beak (meaning a bird’s beak) influences
response time to beaker even though its meaning is not a
component of the meaning of beaker
9/15/10
Psyc / Ling / Comm 525 Fall 10
Eye Movements
• Two types of eye movements
– Smooth pursuit
• Long smooth movements
• Can only do this if eyes are following something
– Saccades
• Short jumps
• Most eye movements
• When eye is not moving = fixations
• Reading consists of saccades and fixations
• Backward saccades = regressions
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
Eyetracking
(Dual Purkinje Tracker)
- Dim infrared light shines on eye
- Reflections bounce back from different layers in eye
- Relative positions of different reflections show where eye is pointing
- Some other kinds of eyetrackers work in different ways
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
Span of Fixation
• How much can you see during a single fixation?
• • It depends on:
–
–
–
–
–
–
your visual acuity
your reading skill level
how hard what you’re reading is, overall
how familiar the current and preceding and next words are
how predictable the current and preceding and next words are
...
– how much of current word you could see on previous fixation
9/15/10
Psyc / Ling / Comm 525 Fall 10
What techniques could be used
to answer this question?
• Eyetracking
– Contingent display changes (McConkie & Rayner)
• Moving window = 1 special type
– Works because you’re functionally blind during saccades
• so you don’t see the change itself happen
– Need eyetracker & computer fast enough to complete display
changes before saccade ends
• Average saccade only lasts 10-20 msec
9/15/10
Psyc / Ling / Comm 525 Fall 10
The person who is fixating where the * is does not see the xxx’s
9/15/10
Psyc / Ling / Comm 525 Fall 10
• So, how far ahead does the eye see?
– Get different answers when
• Ask people what they consciously notice
vs
• See what affects their eye movement patterns
9/15/10
Psyc / Ling / Comm 525 Fall 10
Contingent Display Change
• Idea:
– If something peripheral changes just before the eye reaches it
• during a saccade, so don’t see change itself happening
– If the eye stays on the changed thing longer than when that’s what
was there all along
– what was originally there must have been "seen" peripherally
• Answer to the question:
– You can sometimes get wordshape and initial letter info as far
ahead as 10-14 characters
– But you have to get as close as 6 characters before you detect the
"wordness" of the next unit
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
9/15/10
Psyc / Ling / Comm 525 Fall 10
Semantic Priming
(Phenomenon & Tool)
...
arm
kitchen
tree
Related prime >doctor
nurse
floor
...
<Target>
...
arm
kitchen
tree
actor < Unrelated prime
nurse
floor
...
• In a priming experiment:
– Some people see nurse immediately after doctor in a list of words
= Related condition
– Others see nurse after an unrelated word like actor
= Unrelated condition
- Notice that target word is identical across conditions, so important word
properties like frequency & length are perfectly controlled
– People respond faster, on average, in Related condition
= Priming (= facilitation)
9/15/10
Psyc / Ling / Comm 525 Fall 10