Content uploaded by John Morton
Author content
All content in this area was uploaded by John Morton on Sep 15, 2014
Content may be subject to copyright.
Content uploaded by John Morton
Author content
All content in this area was uploaded by John Morton on Feb 20, 2014
Content may be subject to copyright.
Psychological
Review
1969, Vol.
76, No. 2,
165-178
INTERACTION
OF
INFORMATION
IN
WORD
RECOGNITION
1
JOHN MORTON
Applied
Psychology
Research
Unit,
Cambridge,
England
Quantitative predictions
are
made
from
a
model
for
word recognition.
The
model
has as its
central
feature
a set of
"logogens":
devices
which
accept
in-
formation
relevant
to a
particular word response irrespective
of the
source
of
this
information.
When
more than
a
threshold amount
of
information
has
accumulated
in any
logogen,
that
particular response becomes available
for
responding.
The
model
is
tested against data available
on the
effect
of
word
frequency
on
recognition,
the
effect
of
limiting
the
number
of
response alterna-
tives,
the
interaction
of
stimulus
and
context,
and the
interaction
of
successive
presentations
of
stimuli.
The
implications
of the
underlying
model
are
largely
upheld.
Other possible models
for
word
recognition
are
discussed
as are the
implications
of the
Logogen Model
for
theories
of
memory.
In
previous papers
a
functional
model
for
word
recognition
has
been developed
(Morton,
1964a, 1964b,
1964d;
Morton
&
Broadbent,
1967).
The
form
of
description
used only
lent
itself
to
qualitative
predic-
tions
and
while
it
seemed
to
have some
heuristic
value,
the
overall system
was too
complex
to
allow rigorous specification
of
its
properties.
In the
present paper
the
model
is first
outlined
in a
slightly
simplified
way
and
then
certain
features
of it are
isolated
in
order
to
make
quantitative
predictions about performance
in
word
recognition.
The
various
predictions
made
are
largely independent
and
have
in
com-
mon
only
the
fact
that
in all
situations
there
is
some stimulus information present.
The
effects
of
word
frequency
are
taken
to
indicate relatively permanent changes
in
the
system;
the
effects
of
having
a
reduced
set of
alternative
responses involve tem-
porary changes
in the
same variable.
Different
predictions
are
made concerning
the
interaction
of a
context with
the
stim-
ulus
and the
effects
of
repeated presenta-
tion, these
differences
arising
from
differ-
ences
in the
potential sources
of
such
information.
The
model
contrasts
most
completely with explanations
of
word
1
Part
of
this paper
was
written
while
the
author
was
at the
Department
of
Psychology, Yale Uni-
versity, supported
by
Grant
MH
14229
from
the
National Institutes
of
Mental Health
to
Yale Uni-
versity.
The
author
is
grateful
to W. R.
Garner
for
the use of
much
of his
time.
recognition
which
would
ascribe
all the
observed
effects
as
being
due to
"guessing"
habits.
While
in
conception
the
model
is
very
complex
and
highly interacting,
it
should
be
noted
that
the
separate sections
can be
judged
in
isolation.
In the
description
of
the
model
a
number
of
variables
are
intro-
duced
to
account
for
primary observations.
The
implications
of
most
of
them
are
tested
in
the
sections
that
follow.
DESCRIPTION
OF THE
MODEL
The
basic
unit
in the
model
is
termed
a
logogen.
2
The
logogen
is a
device
which
accepts information
from
the
sensory
analysis mechanisms concerning
the
proper-
ties
of
linguistic stimuli
and
from
context-
producing
mechanisms. When
the
logogen
has
accumulated more than
a
certain
amount
of
information,
a
response
(in the
present case
the
response
of a
single word)
is
made available. Each logogen
is in
effect
defined
by the
information
which
it can
accept
and by the
response
it
makes avail-
able. Relevant
information
can be de-
scribed
as the
members
of the
sets
of
attributes
[&],
\_Vi],
and
[^4»],
these
being
semantic,
visual,
and
acoustic sets, respec-
tively.
More detailed suggestions
as to the
properties
of
these
sets
are
given elsewhere
(Morton, 1968b). Incoming
information
1
From
logos—"word"
and
genus—"birth."
The
author
is
indebted
to
Hallowell Davis
for
suggesting
the
term.
165
166
JOHN
MORTON
Context
System
Semantic
Attributes
Logogen
'<
Available
Responses
Output
Buffer
Rehearsal
Loop
Responses
FIG.
1.
Flow diagram
for the
Logogen Model.
has
only
a
numerical
effect
upon
any
logogen which merely counts
the
number
of
members
of its
denning
sets which occur,
without
regard
to
their origin. When
the
count rises above
a
threshold
value,
the
corresponding
response
is
made
available.
3
Available
responses
go to the
Output
Buffer,
whence
they
may
emerge
as
actual
responses
or be
recirculated
to the
Logogen
System
in a
"rehearsal
loop."
This ideal-
ized
model
is
diagrammed
in
Figure
1.
Since this system operates during reading
and
listening
to
continuous speech,
it is
necessary
to
assume
that
the
value
of the
count decays very rapidly with time,
re-
turning
to its
original value
in
something
of
the
order
of 1
second. Otherwise words
with
a
structural similarity
to the
ones
spoken
would
become available uncon-
trollably. When
a
context
is
presented
in
an
experiment, however,
the
Context Sys-
tem
can
operate almost continuously
to
maintain constant
the
levels
of the
counts
in
logogens
affected
by
that
context.
Stimulus
effects
must remain
transitory
1
It
should
be
noted
that
this
use of the
term
"available
response"
differs
slightly from
the
con-
cept
of
"availability"
as
used
by
other
writers
on
memory (e.g., Tulving
&
Pearlstone,
1966)
and
word
recognition
(e.g.,
Eriksen
&
Browne,
1956).
Pre-
vious
usage
refers
to a
continuum
of
availability
which
corresponds
more
nearly
to the
current
level
of
the
threshold
of a
logogen
in the
present
model.
An
"available
response"
is
related
to the old
term
"implicit
response"
(Miller
&
Dollard, 1941)
but is
more
limited
in its
application
in
some
rather
funda-
mental
ways.
unless
a
portion
of the
stimulus
has
been
recorded
in
some verbalizable
form
such
as
"A
three-syllable word"
or "A
word with
an
initial
p."
Such verbalizations would
act in a way
similar
to
that
of a
context
and
produce
lasting
effects,
which
might hinder
the
subject
if the
information
were
in-
correct.
In the
complete model
the
nature
of the
relationship between
the
Logogen System
and
the
Context System
is
such
that
there
is
continuous exchange
of
information
be-
tween
the
two,
which does
not
result
in
responses becoming available
and
which
is
uncorrelated with
the
objective features
of
the
experiment (see Morton, 1968b).
This
activity
affects
the
values
of the
counts
in
the
various logogens
and it is
assumed
that
samples
of the
values
of the
counts would
be
distributed
in a way
which approximates
the
normal distribution
and
that
all
logo-
gens would have identical distributions.
It is
further assumed
that
such
activity
is
the
only source
of
apparent noise
in the
system. Logogens
can
thus
be
regarded
as
behaving
in a
manner similar
to
detectors
described
by
signal detection theory (Green
&
Swets, 1966). Figure
2
illustrates
the
state
of a
logogen under
various
conditions.
Diagram
2a
represents
its
normal
state,
the
ordinate being
a
probability distribution.
The
range
of the
count
is
such
that
its
value
very rarely exceeds
the
level
of the
thresh-
old
and,
on
average,
t
items
of
relevant
information
are
required
at any one
time
before
the
corresponding response
will
be
available.
The
effect
of a
context
is to
raise
the
mean
of the
count
by an
amount
c,
as in
Diagram
2b. In
this case only
an
average
of
(t—c)
further
units
of
informa-
tion
will
be
required
to
produce
a
response.
The
effect
of a
stimulus
is
similar,
as
shown
in
Diagram
2c,
but,
as
previously men-
tioned,
the
stimulus,
unlike
the
context,
is
not
self-sustaining. When both context
and
stimulus
are
present,
the
units
of
information
add,
and an
average
of
only
[t—(s+c)~}
units
are
required
to
produce
a
response,
as
shown
in
Diagram
2d.
Logogens have
one
further
property,
in
that
following
the
availability
of a
response,
the
threshold
of the
logogen
is
lowered
to a
WORD RECOGNITION
167
certain level
7,
returning
to a
value slightly
less
than
the
original value with
a
time
constant
which
is
very long
in
comparison
with
the
time constant
of the
count decay.
This property
was
originally required
to
explain
the
fact
that
in a
word-recognition
experiment
subjects
often
give
as
erroneous
responses words which have occurred
as
responses earlier
in the
experiment (Morton
1964c).
It is
assumed
that
such
a
property
is
also reflected
in the
fact
that
words
of
high
frequency
of
occurrence are,
in
general,
more intelligible
in
noise
than
low-fre-
quency
words.
In
terms
of the
Logogen
Model
we
would
say
that
logogens cor-
responding
to
words
of
high
frequency
in
the
language have lower thresholds. Since
the
requirement
for
such threshold lowering
is
the
response becoming available,
and not
necessarily
the
response being made, both
frequency
of
emission
and
frequency
of
reception
will
affect
the
threshold.
4
The
short-
and
long-term
effects
of
word repeti-
tion
are
shown
in
Diagram
2e. The
short-
term
effect
is
called
the 7
factor
(and
will
be
a
function
of
time);
the
long-term
effects
are
shown
by the
threshold lines marked
H. F., M.
F.,
and L. F.,
corresponding
to
high-,
medium-,
and
low-frequency
words.
It can be
seen
that
the
effect
of a
lower
threshold
is
equivalent
to the
effect
of
having
contextual
information
in
that
both
reduce
the
amount
of
sensory information
which
would
be
required
to
take
the
level
of
the
count above threshold.
It is
assumed
that
the
system
is
passive,
in
the
sense used
by
Morton
and
Broadbent
(1967),
as
opposed
to
active.
As the
system
operates,
no
comparisons
are
made
by any
mechanism
external
to the
Logogen System
of
the
levels
of
activation
in
different
logogens. Decisions
are
only made within
each logogen. Thus more than
one re-
sponse could
be
available
following
the
presentation
of a
single stimulus.
The
further
assumption
is
made, however,
that
the
exit
from
the
Logogen System
to an
Output
Buffer
is a
single channel,
and
thus
the first
such response
to
become available
will
have
precedence.
*
See
Morton
(1964d)
for a
discussion
of the
emis-
sion
versus
reception
controversy.
Normal
state
Effect
of
context
Effect
of
stimulus
Effect
of
stimulus
plus context
Effects
of
word
frequency
and
word
repitition
FIG.
2. The
effects
of
certain situations upon
the
state
of a
logogen. (The horizontal axes represent
the
level
of
excitation—or
similar
analogy—in
the
logogen.
The
curves correspond
to
probability dis-
tributions
of the
excitation.
The
vertical
lines
represent
the
threshold
of the
logogen. When
the
level
of
excitation exceeds
the
threshold
the
corre-
sponding
word
is
available
as a
response.)
MATHEMATICAL TREATMENT
The
full
treatment
of a
system such
as
the one
described
above
by
signal
detection
theory
has not yet
been worked out.
In-
stead
predictions will
be
made
from
the
response strength model (Luce, 1959).
For
present purposes Luce's model
may be
regarded
as a
logarithmic transformation
of
a
Thurstonian
or a
signal detection
model.
Thus
where
effects
add in the
latter
system,
as in
Figure
1,
they
are
multi-
plied
in the
response strength model.
The
first
step
is to
assign response strengths
to
all
possible responses
in a
given
situation.
This
is
equivalent
in the
Logogen Model
to
assigning
a
value
to the
difference between
the
current level
of
activation
and the
threshold
for
every logogen.
The
proba-
bility
of any
particular response becoming
available
is
then given
by the
ratio
of the
response strength
for
that
item divided
by
the
total
of the
response strengths
for all
the
possible responses.
This
is
termed
the
Ratio
Rule. Since ratios
and not
absolute
differences
are
critical
in
this
form
of
analysis
we are
free
to
scale
the
assigned
values.
Thus
when
we are
considering
the
168
JOHN
MORTON
effects
of a
stimulus
we
will
say
that
the
response strength
of the
correct logogen
is
a
compared with
a
value
of
unity
for all
other logogens.
This
is not to say
that
the
stimulus
has no
effect
on the
other logogens,
merely
that
on
average
the
effects
will
be
the
same
for all the
others.
The
value
of a
then
is
properly regarded
as the
difference
between
the
effects
on the
stimulus logogen
and the
average
of all
others.
The
state
of
the
logogens when
the
stimulus
is
presented
is
contrasted symbolically with
the
effect
of
the
stimulus
by use of the
letter
V.
This
letter
will
be
used
to
designate
the
differ-
ences
in
threshold
between logogens
and
also
the
differential
effects
of
context.
The
combined
effects
of the
stimulus
to-
gether
with
one of the
other
factors
is
calculated
by
multiplying
the
response
strengths
of the two
effects
for
each logogen.
This
is
equivalent
to
adding
the
effects
in
Figure
Id.
In the
case where
we
have
assigned
a
value
of 1, as in
Table
2, the
multiplication
in the
response
strength
analysis
is
equivalent
to
saying
that
there
is
no
effect
of the
stimulus
on
that
logogen.
Throughout this paper
it
should
be re-
membered
that
the
mathematics
is
only
used
to
predict
the
operation
of the
system
which
is
described above.
PREDICTIONS FROM
THE
MODEL
Word-Frequency
Threshold
Effect
A
number
of
authors have shown
that
intelligibility
in
noise
and the
visual dura-
tion threshold
for
words
are
strongly
in-
fluenced by the
frequency
of
occurrence
of
the
words,
and
there
has
been much con-
troversy
as to
whether such results should
be
attributed
to
"stimulus effects"
or
"response effects" (Broadbent, 1967).
Within
the
Logogen Model, which
in
this
respect resembles Broadbent's
own
treat-
ment,
the
dichotomy
is
scarcely applicable
(Morton, 1968a).
By the
model
the
difference
between
words
of
differing
frequency
of
occurrence
is
that
the
logogens
have
different thresh-
olds. Thus
the
logogens corresponding
to
high-frequency
words
will
require less
stimulus information
for the
count
to
rise
above
the
threshold.
The
amount
of
stimulus
information available
to a
logogen
is,
however, assumed
to be
independent
of
the
frequency
of the
word
and is a
function
only
of the
properties
of the
stimulus
(duration,
contrast, signal-to-noise ratio).
Thus
we
will
say
that
the
presentation
of a
stimulus
increases
the
response strength
of
the
stimulus logogen
to an
amount
a
com-
pared with
all
other logogens, whose
strength remains
at
unity. (Since
we are
concerned with ratios,
the
latter
value
is
arbitrary
and
simply serves
to
scale
a.)
This
is not to say
that
it is
assumed
that
all
other
words
are
equally confusable with
any
stimulus
word, merely
that
stimulus con-
fusions
are
unrelated
to
word frequency.
This
is the
"principle
of
acoustical
equi-
valence" suggested
by
Brown
and
Ruben-
stein (1961). These authors investigated
the
word-frequency threshold
effect,
divid-
ing
the
6500 monosyllabic content words
into
13
classes
of 500
words each,
the
classes
being formed
by
grouping
the
words
by
their frequency.
They
presented sub-
jects (5s) with
a
total
of
1300 words, con-
sisting
of a
randomized selection
of 100
words
from
each
frequency
interval
in
noise,
and
calculated
Ci, the
number
of
responses which were correct
in
each inter-
val,
i, and
s,-,
the
number
of
responses which
were
in the
same frequency interval
as the
stimulus
(but might have been incorrect
responses).
8
The
latter
measure includes
d,
however,
and
so
cannot
be
compared with
it;
instead
we
will
use
e
it
the
number
of
incorrect
re-
sponses
in the
same
frequency
interval
as
the
stimulus.
Morton
(1968a)
has
tested
a
series
of
models
against Brown
and
Rubenstein's
data,
rejecting single-threshold models
which claim
that
the
word-frequency
effect
is
entirely
due to
response bias
and a
model
by
which
the
effect
is
entirely
due to the
stimulus,
in
which
a
would
be a
function
of
word
frequency.
For the
model
at
present
under
discussion,
the
response strengths
appropriate
to the
different
responses
in
this situation
are
given
in
Table
1. If the
6
The
author
is
grateful
to H.
Rubenstein
for
kindly
providing
his
original
data.
WORD RECOGNITION
169
TABLE
1
RESPONSE STRENGTHS
IN THE
WORD-
FREQUENCY
EXPERIMENT
Stimu-
class
1
»
n
Correct
aponses
aiVi
aiVi
ctnV*
Incorrect
responses
of
interval
1
(M-l)Fi
MVi
MVi
i
MV
f
(M-l)Vi
MVi
n
MV
n
UVn
(M-DV,
Total
of
strengths
T,
Ti
r
n
assumption
concerning
the
invariance
of the
amount
of
stimulus information with word
frequency
is
correct,
all a
will
be
equal.
The
differences
in
thresholds
of
logogens
in
different frequency
classes
are
indicated
by the
variables
V\-
• •
W
• •
V
n
.
M is the
number
of
responses
in
each
frequency
class.
The
probability
of a
correct
re-
sponse
in the
interval
i is
thus
given
by the
equation
^•
=
^7'.
[1]
The
probability
of an
error
response
being
in the
correct
frequency
interval
is
given
by
«i
=
r
'•
C
2
]
Combining
Equations
1 and 2 we
obtain
c
t
=
eg
M-l
[3]
Since
in the
present model
a,
is
independent
of
i, we
predict
a
linear relationship
be-
tween
d
and
e
f
.
The
data
are
plotted
in
Figure
3
with
the
best
fitting
straight
line.
The
nonzero
intercept
and the
possible
slight
curvature
in the
data
are
capable
of
several explanations,
any of
which would
require
slight
modifications
to the
model.
However,
all the
alternative
models
tested
can be
rejected
out of
hand
(Morton,
1968a).
The
Interaction
of
Set
Size with Signal
Level
A
number
of
experiments have been per-
formed
in
which
5s
have
been
presented
with
a
stimulus word which
has
been
selected
from
a
restricted
set of
known
alternatives.
The
general result
is
that
the
intelligibility
of a
stimulus
is
increased
as
the
number
of
response
alternatives
is
reduced. With
the
present
model
the
known
alternatives
would have
their
thresh-
olds
lowered
to a
level
7.
This
would
serve
to
make
the
response
strengths
of
other
responses small
by
comparison;
to a first
approximation they
can
then
be
ignored
in
the
analysis.
We can
then regard
the
response strength
of the
stimulus word
as
being
a
compared with unity
for all the
other alternatives.
This
again assumes
equal
confusability between
the
alterna-
tives—an
assumption which
is
certainly
not
true
and
would result
in a
relative lowering
of
performance.
In
addition
it
assumes
that
the
average confusability
of the
alternatives
is
independent
of
their number.
Since,
as
before,
a
represents
the
difference
between
the
stimulus information received
by the
stimulus logogen
and the
average
of
that
received
by
other logogens, violation
of
this assumption would lead
to
different
S/N = Odb
0123
Probability
of an
incorrect
response
in
the
correct
frequency
class
(e,l
FIG.
3.
Data
from
Brown
and
Rubenstein
(1961).
(The
abscissa,
et,
is the
probability
of an
incorrect
response
to a
stimulus
in the
*th
frequency
interval
itself
being
in the
ith
interval.
C, is the
probability
of
a
response
to a
stimulus
in the
»th
interval
being
correct.
The
line
is the
least squares
fit.—From
Morton,
1968a.)
170
JOHN
MORTON
O.98
0.95
50.80
050
0.02
2 4 8
16
32 296
1000
Number
of
Alternatives
FIG.
4.
Data
from
Miller,
Heise,
and
Lichten
(1951),
showing
variation
in
performance
with
dif-
ferent
numbers
of
alternatives.
(The
function
plotted
is:
logit
P,
= log a -
log
(N
- 1), The
lines
are the
least squares
fit.
Logit
P
n
= log
predictions. While
it is
unlikely
to be
true,
it
will,
however, serve
as a first
approximation.
With
these
assumptions
the
probability
of
producing
a
correct response
at any S/N
with
N
alternatives
is
given
by:
[4]
±n
a+(N-lY
This
can be
rewritten
as
P»
1-P.
N-l
whence
[5]
[6]
where
logit
P
n
=
log[P
n
/(l-P
n
)].
Miller,
Heise,
and
Lichten (1951) provide
data
against which this equation
can be
tested.
In
their experiment
5s
were
pre-
sented with spoken monosyllables
in
differ-
ent
levels
of
masking noise.
The
words
were
chosen
from
known vocabularies
of
sizes
ranging from
2 to
1000.
The
functions
resulting
from
entering their
data
into
Equation
6 are
plotted
in
Figure
4.
Only
those
S/N are
used
for
which
four
or
more
points
are
available.
It can be
seen
that
the
deviations
from
linearity
are
well
within
the
experimental error
in the
data.
The
average slope
of the
resulting lines,
fitted
by the
method
of
least squares,
is — .82
which
is
reasonably close
to the
predicted
slope
of
minus one.
There
are
several well-motivated
modifi-
cations
of the
model which
would
lead
to
different
predictions,
but
attempts
to
apply
these
would
only
add a
small amount
to the
predictive
and
heuristic value
of the
model.
The
following
are
examples
of the
kinds
of
modification
which
might
be
made:
1.
Breakdown
of the
assumption
of
equal
average
confusability
of the
alternatives
could
lead
to
relative changes
in
per-
formance
at any
value
of N in
either
direction.
2.
With
a
small number
of
alternatives
it is
possible
that
5s
would
try to
predict
in
advance
the
next item
in the
sequence.
As
the
stimuli were randomized, this
attempted
prediction could only lead
to an
overall worsening
in
performance.
3.
With larger numbers
of
alternatives
it
is
possible
that
5s
were
in
fact
using
a
smaller
number than
that
prescribed.
The
likelihood
of
their recognizing stimuli
out-
side
the
subjective
set
would
be
much lower,
but for
values
of a
which were small com-
pared with
N it is
possible
that
performance
would
improve.
On the
other hand,
if the
subjective
set
were larger than
the
objective
set—if,
for
example,
in the
1000-word con-
dition
the
subjective
set
were
the
complete
set of
English
monosyllables—performance
on
that
condition
would
be
apparently
worse.
4.
The
present model assumes
that
the
distribution
of
"noise"
in the
logogens
is
logistic.
6
Different
assumptions
about
the
distribution, equally
well
motivated, could
make
the
model
fit
better
or
worse.
5. If the
response strengths
of the
items
outside
the
objective
set are not
negligible,
we
would obtain
the
equation
logitP
B
=
loga-log[(^-l)+fe]
[7]
where
k is the sum of the
response strengths
of
the
other items. Introduction
of
this
8
The
logistic
distribution,
required
by the Re-
sponse
Strength mathematics,
is
similar
to the
normal
distribution,
differing
chiefly
in the
tails.
WORD
RECOGNITION
171
constant
could make
the
model
fit the
data
almost perfectly.
None
of
these
modifications
would
affect
the
underlying principles
of the
model.
The
best
we can say
then
is
that
the
model
fits the
data
reasonably
well.
This
is as
much
as any
model
can
hope
to do. It is,
however, clear
that
by the
above analysis
the
effect
of
increasing
the
signal strength
is
the
same regardless
of the
number
of
alternatives.
Alternative
Models
1.
Green
and
Birdsall
(1958)
have plotted
Miller
et
al.'s
data
by
calculating values
of
d'
for the
different
values
of N,
showing
that
d'
is
approximately independent
of N.
This
procedure
closely
approximates
the
one
used above
and has
essentially
the
same
underlying
model.
2.
Garner
(1962)
has
plotted
the
same
data
in
terms
of the
number
of
bits
of
information
transmitted, showing
that
this
measure
is
essentially independent
of
TV.
This
form
of
analysis makes
no
claims about
the
form
of the
underlying model
and is
perhaps
better
regarded
as
indicating,
in
general,
that
5s'
efficiency
remains inde-
pendent
of N. As
such
it is not an
alterna-
tive
model
to the
Logogen Model,
but
rather
an
alternative method
of
treating
the
data.
3.
Stowe, Harris,
and
Hampton (1963)
propose
a
model whereby words
are
dis-
criminated
by
identifying them
on a
num-
ber of
binary stimulus dimensions.
The
correct response
is
only given
following
correct
identification
on all the
dimensions.
Logs
n
dimensions would
be
required
to
discriminate among
n
words. Thus
in the
two-choice situation only
a
single dimension
would
be
used.
If the
probability
of a
correct
decision
on one
dimension were
Pd,
and if all
dimensions were equally well
discriminated, then
the
probability
of a
correct
identification among
n
alternatives
would
be
given
by
P
n
=
(P<0
log2B
.
This
model
gives
a
reasonable
fit to
Miller
et
al.'s
data.
It
appears
to
suffer
a
conceptual
drawback, however,
in
that
as the
number
of
alternatives
is
reduced,
the
number
of
discriminating dimensions which
the
system
uses
is
reduced. Some
of the
dimensions
excluded
from
the
discrimination
process
(presumably
in
some
arbitrary
way) could
in
fact provide usable information.
The
model
is
insufficiently
explicit
to
enable
further
discussion.
It has
been suggested
that
the
effect
of
restricting
the
number
of
alternatives
is to
reduce
the
thresholds
of the
appropriate
logogens
to a
common value
7. If
this
were
so,
then
we
would expect,
as
Pollack,
Rubenstein,
and
Decker (1959) have shown,
that
word frequency
does
not
exert
a
measurable
influence
upon
the
intelligi-
bility
of
words
in
known message
sets.
The
Interaction
of
Stimulus
and
Context
There
is a
large
body
of
evidence
which
indicates
that
the
recognition
of a
word
is
greatly
facilitated
by the
prior presentation
of
a
context.
The
extent
of the
facilitation
is
a
function
of the
likelihood
of the
context
eliciting
the
word
in a
free-response
situ-
ation.
In
most cases
the
context
consists
of
an
incomplete sentence which
the
stimu-
lus
word completes,
but the
analysis below
can
be
generalized
to any
context.
Three
sources
of
data
are
required:
the
proba-
bility
of the
target
word being given
as a
response
to the
context alone,
in the
absence
of a
stimulus;
the
probability
of the
response
to a
stimulus
at a
particular
S/N
in
the
absence
of a
context;
and the
probability
of the
response
at
that
S/N
when
the
context
has
already
been
pre-
sented.
Table
2
gives
the
appropriate
response
TABLE
2
RESPONSE
STRENGTHS
WITH
STIMULUS,
WITH
CONTEXT,
AND
WITH
BOTH
Experimental
condition
Stimulus
only
Context
only
Stimulus
and
context
Correct
re-
sponse
a
Vt
>F<
Other
responses
to
individual
words
1
!•••!
ViVi-V.
Fl
Vf'Vn
Total
of
response
strengths
j
*
B=a
2-i
'
i
T+(
a
-l)V
t
Note.—The
entries
in the
third
line
are
obtained
by
multi-
prying
the
entries
in the first two
lines.
This
is
equivalent
to
adding
the
effects
of
stimulus
and
context
in
Figure
2d.
172
JOHN
MORTON
strengths
for
stimulus alone, context alone,
and for the
combination
of
stimulus
and
context.
The
response strengths
for the
combined condition
are
obtained
by
multi-
plying
the
strengths
in the
other
two
conditions.
This
table
is
intended,
as
before,
to
represent
the
average
of all
possible outcomes.
The
probability
of a
correct response
from
the
stimulus,
P
a
,
alone
is
given
by the
equation
^
M
[9]
The
probability
of a
correct response
from
context alone,
P
0
,
is
given
by
from
which
we
obtain
from
which
.
T
V
t
/P
t
.
[10]
o
8
word context
a
4
word
context
A
0
word
context
20
40 60 80
100
Exposure
Duration
in
milliseconds
140
FIG.
5.
Data
from
Tulving,
Mandler,
and
Baumal
(1964)
on the
joint
effects
of
stimulus
and
context. (The
function
plotted
is:
logit
P,
c
=
a,
+ bx,
where
x is the
exposure duration,
P,
e
is the
probability
of a
correct response,
and a and b are
constants.)
With both context
and
stimulus,
the
probability
of a
correct response,
P
sc
,
is
«V<
[12]
If
we
substitute
for a and T
from
Equations
9
and 11 in
this equation,
F,-
vanishes
and
we
are
left,
after rearrangement,
with:
P..
P,
1-P..
1-P.
I-Pc
•(N-l).
[13]
If
we
take
logarithms this equation
can be
written
as
:
logitP
sc
= l
H-log(TV-l).
[14]
It
should
be
noted
that
the
term
log
(N — 1) is not a
variable
which
can be
considered
as
varying
in the
presence
of a
context, with
or
without
a
stimulus, since
it is
derived entirely
from
the
condition
where only
a
stimulus
is
given.
It is
thus
unaffected
by any
effect
the
context
may
have
in
reducing
the
number
of
alternative
responses.
In
fact, Equation
14 can be
written
as
:
logitP
8C
=
loga+logitP
c
.
[15]
Tulving, Mandler,
and
Baumal (1964)
provide
data
against
which
this
equation
may
be
tested.
They presented
5s
with
18
words
at
successive durations ranging
from
zero,
in
fact
a
"word-like smudge,"
to 140
milliseconds.
The 5s had
addi-
tional information
in the
form
of
zero, two,
four,
or
eight words
of a
context sentence.
They propose
an
empirically derived equa-
tion
for
their
data
:
logitP«
=
[16]
where
a,-
is a
constant depending upon
the
level
of the
context,
&
is a
constant,
and x is
the
duration
of
exposure
of the
stimulus
word.
They
do not
present their
data
in
the
form
of
this equation,
so it is
given
in
Figure
5
with lines drawn
from
the
con-
stants
they
give.
The
data
for
two-words
context
are not
given
as
they
are
almost
identical
to the
values
for
four-words
context.
The
data
are a
moderately good
WORD RECOGNITION
173
fit, but
there
is a
curvilinear component
apparent
in all
three
of the
context con-
ditions.
In
Figure
6
these
data
are
replotted
according
to
Equation
14
with
the
context
level
as the
parameter.
This
has the
effect
of
scaling
the
abscissa (the duration)
according
to the
performance
in the
zero
context
condition. Lines
of the
predicted
slope,
unity, have been
fitted by
eye,
and
it
will
be
seen
that
they approximate
the
data
excellently except
for the two
highest
exposure durations
for
eight-words context.
From
this
we may
conclude
that,
within
the
model
we are
using,
the
amount
of
stimulus
information
available
to the
Logogen
System does
not
change when
a
context
is
present
and
does
not
vary with
the
amount
of
context information.
In one
sense
stimulus
information
is
independent
of
context information. Such
a
statement
is
only
meaningful
within
the
framework
of a
particular model
and
should
not be
inter-
preted
in
terms
of the
statistical
inde-
pendence
model
1
sc—-t
a\~±
c
-L
B-L
c
[17]
which
is
logically
inappropriate
in
this
situation
since
it
would imply
that
the
sources
of
information were equivalent.
Clearly
S
1
would
not
choose
his
preferred
"correct"
response
to the
context
in the
face
of
apparently contradictory
informa-
tion
from
the
stimulus
as
would
be
implied
by
Equation
17.
Equation
14
also predicts
that
logit
P
ac
will
be
linear with logit
P
c
with
a
slope
of
one.
In the
case
of
Tulving
et
al.'s
data
we
would
have
only
three
points
per
line.
In
addition
the
estimate
for
P
c
is
extremely
unreliable,
consisting
of
only three correct
responses
out of a
possible 450.
The
differences
between
P
ac
with
four
and
eight
words
of
context
are of the
right order
of
magnitude
at all
except
the two
highest
durations,
the
mean increase
in
logit
P
80
being
.332
for an
increase
in
logit
P
c
of
.325.
There
do not
appear
to be any
further
data
available
against
which
to
test
Equa-
tion
14.
Many authors have investigated
the
effect
of
context upon
recognition,
but
.05
.1
.2 .3 A S .6 .7 .8
Probability
correct
with
stimulus
alone
(P
s
)
FIG.
6. The
same data
as in
Figure
S;
plotted
as
logit
P,
e
=
K
+
logit
P..
(The lines
are fitted by
eye
with
the
predicted slope
of
unity.)
the
form
of the
experiments renders them
unsuitable,
mainly because
the
method
of
ascending limits
or
some similar procedure
has
been used
for
each individual stimulus.
In
the
earlier discussion
it was
pointed
out
that
the
effect
of a
response becoming
available
was to
lower
the
threshold
of the
logogen
to a
level
y.
Such
an
effect
would
occur
whether
the
response were called
for
by the
experimental procedure
or
resulted
from
S
actively attempting
to
complete
the
sentence
(or
fulfill
the
requirements
of
another
context)
prior
to the
presentation
of
the
stimulus.
If the
prior response
coincided with
the
stimulus
we
would
expect
an
increase
in the
proportion
of the
correct
responses
;
if the
prior response
did
not
coincide with
the
stimulus
we
would
expect poorer performance.
The
state
of
the
response-strength
table
following
vari-
ous
outcomes
is
given
in
Table
3.
Where
the
correct response
has
been given prior
to
the
presentation
of the
stimulus (what
we
may
call
the
7-plus
case)
the
subsequent
performance
is
predicted
by the
equation
logitP..
=
loga
+
logitP
c
+
[18]
The
7-minus
case, where
a
response
has
been
given which
is
different
from
the
stimulus,
is not
amenable
to
solution
with-
out
knowing
the
complete distribution
of
responses
to the
context.
The
full
solution
174
JOHN
MORTON
TABLE
3
RESPONSE
STRENGTHS
WHEN
AN
ANTICIPATORY
RESPONSE
HAS
BEEN
GIVEN
Situation
No
prior response
Correct prior response
(-v-plus)
Incorrect prior response (7-minus)
Correct
response
aVi
ctyVt
aVi
Incorrect
responses
y
V
•
«
•
V
••
.
.
V
ViVf-Vj---V
n
7,7,".-yVy...7.
Total
r+(a-l)7i
T + (ay — 1) Vi
T
+
(a-
1)7,.
+
(7-DF,-
Note.—r
=•
2 Vi
throughout.
The
effects
on the
response strength
of a
prior response
are
obtained
by
multiplying
the
appro-
priate entry
by 7.
This
is
equivalent
to a
reduction
in
threshold
of an
amount
7 in
Figure
2e.
has the
form
p
_
86
0V*)
where
the
values
of
P
s
represent
the
prob-
abilities
of
words other than
the
target
word
(Word
t)
being elicited
by the
con-
text,
and
P,
c
is the
average outcome
for any
one
target
word.
An
approximation based
on
the
assumption
that
all
other responses
have equal response strengths gives
the
equation,
to be
compared with Equation
15,
logitP
8C
=
loga
+ log
This
equation would underestimate
the
effects
of 7,
especially
for low
probability
words.
Rubenstein
and
Pollack (1963) provide
data
against
which some
of the
implications
of
these equations
can be
tested.
They
presented
their
Ss
with
a
context
:
an in-
complete sentence,
a
word
to be
associated
to,
a
category name,
or the first 1-5
letters
of
the
target
word. Each
5
made
a re-
TABLE
4
DATA
FROM
RUBENSTEIN
AND
POLLACK
(1964)
FOR
A
SENTENCE CONTEXT WITH
S/N
OF
-
14db.
Group
Performance
p.
p..
7-plus
(R)
7-minus
(F)
.02
1.00
.02
.11
.92
.10
.20
.70
.15
.26
.99
.14
.35
.97
.20
.47
.94
.24
.68
.97
.25
.84
.95
.37
Note.—Data
given were
the
overall probability correct (Q),
the
probability correct
for the
7-minus group
(F),
and
Po.
The
7-plus
score
(£)
was
estimated
from
Q
=
P,-E
+ (1 -
P.)F,
whence
E
-
CQ
- (1 -
j
sponse
to the
context ("the word they
thought most likely
to
occur")
and
then
was
presented with
the
stimulus word
at
six
increasingly favorable S/N, making
a
response
following
each presentation.
Rubenstein
and
Pollack present
data
in two
forms,
the
average proportion
of
correct
responses
for the
whole group
and the
proportion
for
those
5s who
made
an
initially
"incorrect"
response
to a
particular
context, which
we may
call
the
-/-minus
group. When
the
performance
for the
7-plus group
is
estimated
from
the
data
it
is
apparent
that
they made very
few
errors
even
at the
lowest S/N,
as
would
be ex-
pected
from
the
above analysis.
The
performance
of the
7-minus group
is
con-
siderably worse,
and at the
lowest
S/N
their performance
is
worse than
that
of the
whole
group prior
to the first
stimulus
presentation.
The
relevant
data
are
given
in
Table
4.
From Equation
20 it
seems likely
that,
with
a few
more assumptions concerning
the
effects
of
sequential presentation,
the
effects
of
increasing
the S/N
will
be
inde-
pendent
of the
value
of
P
e
,
provided
that
the set of
possible responses remains
approximately orthogonal with regard
to
their stimulus properties.
The
results
for
the
sentence context
are
presented
in
Figure
7. The
ordinate represents
the
difference
in
logit
P
sc
between
the
—
14db.
condition
and the
parameter
level;
that
is,
effectively,
the
result
of
increasing
the
signal level
from
—14
db. to the
given level.
It is
apparent
that
there
is no
systematic
change.
In
contrast
the
data
for the
letter context
are
given.
In
this condition
the
context provides information which
is
highly
correlated with
the
information
WORD
RECOGNITION
175
available
in the
stimulus.
In
other
words
the
resulting
set of
possible response words
will
be
more highly
confusable
on
average,
the
larger
the
number
of
letters
context.
Thus
we
would predict
that
the
value
of a
would
fall,
the
greater
the
number
of
con-
text
letters.
This
follows
from
the
opera-
tional
definition
of a as the
difference
be-
tween
the
amount
of
information available
from
the
stimulus
to the
correct logogen
compared with
the
average
for all
other
relevant logogens. This predicted trend
is
clear
in the
data.
Interaction
of
Successive Presentations
One
property
of the
Logogen
System
which
has not
been discussed
is the
sugges-
tion
that
stimulus
information
is
only
effective
for a
relatively
short
period after
presentation. Thus
we
would
expect
no
interaction between successive presenta-
tions
of the
stimulus, unless
they
followed
in
very quick succession. Various studies
may be
cited
in
support
of
this contention.
Postman
and
Adis-Castro (1957) com-
pared performance
on
word recognition
with
the
method
of
ascending limits
and the
method
of
random series whereby
all the
words were presented
for
recognition
at
each duration
before
proceeding
to the
next
level.
They
found
no
significant
difference
between
the
conditions. Pollack (1964)
investigated
the
interaction
of
visual
and
auditory information
in
word
identification.
He
discovered
that
successive presentations
of
a
stimulus word visually
and
auditorily
gave rise
to
only
2%
more correct responses
on
the
second modality than
would
have
been
expected
from
the
statistical
inde-
pendence model.
He
further
points
out
that
if
partially correct responses were
credited,
even
this
2%
difference
would
vanish. This result
effectively
says
that
if
5
identified
the
word
from
the first
modal-
ity,
he
identified
it
from
the
second
one
(owing
to the
operation
of the
y-effect);
if
he did not
recognize
the
word
on the first
modality,
the
probability
of a
correct
identification
on the
second modality
was
the
same
as it
would have been
if the first
modality
had not
been presented.
Letter
Context
234567
.11
.20
.26
.35 .47
.68
2345
I
I I I I
.02
.17
.46 .73 .79
P
c
FIG.
7.
Data
from
Rubenstein
and
Pollack
(1963),
showing
the
difference
in the
value
of
logit
P,
c
between
the
parameter
S/N and —
14db.
for
two
kinds
of
contexts. (With
the
sentence context
the
effect
of
increasing
the
S/N
is
independent
of the
extent
to
which
the
context predicts
the
stimulus
word.
With
a
context
of the first 1-5
letters,
the
effect
of the
stimulus
is
less
the
more
the
context
owing
to the
correlation between contextual
and
stimulus
information.)
Pollack's technique
was to
present
the
stimulus
word once
in one
modality
and
then
six
times
in the
other modality under
increasingly favorable conditions.
He re-
marks
that
"Control
tests
were carried
out
to
ensure
that
the
successive ascending
procedure
was
unbiased with respect
to
independent
presentations
[1964,
p.
78],"
essentially
confirming
Postman
and
Adis
Castro,
and
while
he
quotes
no
data
to
support
the
statement,
the
main result
is
justification
in
itself.
Now
it is
clear
to
anyone
who has
examined
the
successive responses
of in-
dividual
5s in an
experiment using
the
ascending
method
of
limits
that
there
are
dependencies
in
successive responses, par-
ticularly
in
that
incorrect responses
often
influence
subsequent responses (Morton,
1964a).
It is
also clear
to
anyone
who has
been
an 5
under such
a
procedure
that
if,
for
example,
on the first
presentation
the
beginning
of the
word
has
been seen clearly,
one
can
profitably
concentrate
attention
on
the end of the
word
on the
next presenta-
176
JOHN
MORTON
tion.
What
must
not be
forgotten,
how-
ever,
is
that
such partial evidence
can be
incorrect
as
well
as
correct.
If it is in-
correct, then subsequent recognition
will
be
hampered,
as
apparently occurred with
Rubenstein
and
Pollack's
y-minus
5s.
What
the
experimental
data
show, then,
is
that
these
two
competing
effects
tend
to
balance
out
over
the
course
of an ex-
periment. From this analysis
we
would
predict
that
if 5s
were
informed
whether
or
not
their
partial
responses were
correct
their performance should improve
on
sub-
sequent trials, since,
it can be
assumed,
the
incorrect
information, sustained
as it is in
the
present model
by
some system equi-
valent
to the
Context
System, would
not
be
passed
to the
Logogen System. Such
a
result
was
found
by
Pollack, Rubenstein,
and
Decker (1959).
In
summary,
it is
apparent
that
a
com-
plete description
of the
processes contri-
buting
to the
effects
of
repeated presenta-
tion
and the
prediction
of
performance
will
require
detailed consideration
of
partial
responses
as
well
as
full
responses. Gross
probabilistic predictions ignore elements
of
the
experimental situation which
are
clearly
important.
CONCLUSIONS
The
model under discussion
was
origin-
ally devised
for
qualitative
reasons;
in
this
paper
quantitative
predictions have been
made about
5s'
behavior
in a
variety
of
situations. Although
the
model
is
ideal-
ized
in
ways which have been
discussed,
the
data
do not
seem
to
contradict
it. At the
moment
there
is a
lack
of
alternative
func-
tional models against which
the
Logogen
Model
can be
tested.
Purely mathematical
models have been ignored because,
as we
have seen,
it is
relatively simple
to add an
extra parameter,
albeit
intuitively justified
as a
variable, which
can
produce
a
perfect
fit to the
data.
Such
a
practice
has not
been
indulged
in
since there
are
several
factors
with equal claim
for
inclusion whose
quantitative
effect
cannot
at the
moment
be
estimated outside
the
data
they would
be
accounting for.
The
existing model
has
much
in
common
with
one
recently
put
forward
by
Norman
(1968).
One
difference
hinges
on the re-
spective
treatments
of
memory. Norman
first
points
out
that
it
seems necessary
to
distinguish between
two
forms
of
memory
which
correspond, respectively,
to our
immediate memories
of
events
and
memor-
ies
of
events
a few
seconds old. Norman
follows
William James
in
calling these
types
"primary"
and
"secondary,"
but
makes
the
unnecessary assumption
that
the
primary memory must precede secondary
memory
in the
processing system,
an
assumption carried over
from
Waugh
and
Norman
(1965).
In the
recent paper
Norman
argues
that
"there
must
be
sufficient
interconnections between
the
storages
to
allow
a
comparison
of the
just-
perceived sensory events with
the
collection
of
previously experienced perceptions
[pp.
524-525]".
In
fact, then,
the two
systems
become
effectively
a
single system with
two
types
of
storage within
it.
The
Logogen Model does
not
suffer
from
these objections
for we can say
that
the
Primary Memory
is
located
in the
Output
Buffer
which
follows
the
Logogen System.
Traces
remaining
in
logogens, either
in the
values
of the
count
or the
threshold, could
then
be
regarded
as two
possible sources
of
information
(with
different
time charac-
teristics)
for
Secondary Memory.
The
properties
of
these
two
stores, derived
as
they
are for
reasons other than memoric,
match
fairly
well
the
usual characteriza-
tions
of
Primary
and
Secondary Memory.
The
Output
Buffer,
which
manifests
itself
in the
eye-voice span
(Morton,
1964b)
and the
ear-voice
span (Treisman
&
Geffin,
1967)
is
seen
as
having
a
limited
capacity,
as is
primary memory. Material
within
it is
coded
in
terms
of
articulation
parameters
(since
the
usual
destination
of
the
material
is
speech). Thus
we
would
expect
to find
articulatory
confusions
in
Immediate Recall
for
visual stimuli,
which
task
may be
taken
as
involving primary
memory.
Conrad (1964)
has
shown
that
errors
in
recall
of
visually presented letters
are
highly correlated with errors made
in
recognition
of
letter
names presented audi-
WORD
RECOGNITION
177
torily
in
noise. Conrad terms these
"acous-
tic
confusions"
but the
correlation between
acoustic
and
articulatory descriptions
leaves
open
the
possibility
that
the
correct
description
is
"articulatory confusions."
The
Logogen System,
on the
other hand,
like
secondary memory,
is not
subject
to
overwriting
insofar
as the
7-effect
is
con-
cerned.
It
would, however,
be
impossible,
from
the
state
of the
appropriate logogen,
to
discover whether
or not a
stimulus
had
been presented once
or
twice. Crowder
(in
press)
has
shown
that
lists
of
letters
or
words containing
repeated
items
are
more
difficult
than those without repeats. Such
information
as is
retrieved concerning
repeats
would
be
available either
in the
Output
Buffer
or in a
long-term memory
store.
The
latter
construct
is
necessary
to
account
for our
ability
to
maintain
the
effect
of a
context,
and
could
well
account
for
such
associative
phenomena
as are
demonstrated
in
memory
tasks.
It is of
interest
that
the
present model,
derived
to
account
for
phenomena
of
word
recognition, requires
at
least three con-
structs,
each with
different
properties,
from
which
information
could
be
retrieved
in a
memory
experiment. Arguments based
on
parsimony
as to
whether there
are one or
two
separate memory stores (Atkinson
&
Schiffrin,
1967;
Melton, 1963)
tend
to
lose
their
force.
It
might
be
more profitable
for
future
work
on
memory
to
concentrate
on
attempting
to
specify
the
relative con-
tributions
of
these
different
information
sources
in
different
experimental
conditions.
7
The
differences
in
detailed treatment
of
memory
between Norman's model
and the
Logogen Model
are
almost trivial
and are
certainly reconcilable.
Of
more interest
are the
similarities between
the
models.
The
core
of
Norman's model
is a
storage
system into
which
sensory inputs
and
"pertinence"
inputs
are
sent.
The
ele-
ments
of the
storage system
are
almost
identical
to
logogens
in
their properties.
Further,
the
sensory analysis systems
in
both models
are
conceived
of as
passive,
or
autonomous. Although Norman refers
to
7
R. G.
Crowder
and J.
Morton. Precategorical
acoustic
storage.
Unpublished manuscript, 1968.
his
model
as a
modified
analysis-by-
synthesis
model,
the
synthesis
is in
terms
of
"expectations,"
and as
such
takes
a
differ-
ent
form
from
other
analysis-by-synthesis
systems which require
that
the
synthesized
anticipation
is in the
same code
as the
input. Norman's model also provides
a
framework
within which problems
of
atten-
tion
and the
retrieval
of
information
can be
discussed, neither
of
which topics
the
Logogen Model
is
capable
of
handling
in its
present
form.
Thus
the two
models
can
be
seen
as
complementary
to one
another.
Perhaps
the
most
important
similarity
is
a
strategic
one. Both
the
models
are
capable
of
expansion
to
deal with phenom-
ena
outside
the
areas
for
which
they
were
evolved without losing their essential
character
and
without becoming
too
rigid.
REFERENCES
ATKINSON,
R.
T.,
&
SHIFFRIN,
R. M.
Human
memory:
A
proposed system
and its
control proc-
esses.
Technical
Report
No.
110,
Institute
for
Mathematical
Studies
in the
Social Sciences,
Stanford
University, 1967.
BROADBENT,
D. E.
Word frequency
effect
and re-
sponse
bias.
Psychological
Review,
1967,
74,
1-15.
BROWN,
C.
R.,
&
RUBENSTEIN,
H.
Test
of
response
bias explanation
of
word-frequency
effect.
Science,
1961,
133,
280-281.
CONRAD,
R.
Acoustic
confusions
in
immediate
memory. British
Journal
of
Psychology,
1964,
55,
75-84.
CROWDER,
R. G.
Intraserial repetition
effects
in
immediate memory. Journal
of
Verbal
Learning
and
Verbal
Behavior,
in
press.
ERIKSEN,
C. W., &
BROWNE,
C. T. An
experi-
mental
and
theoretical
analysis
of
perceptual
de-
fence.
Journal
of
Abnormal
and
Social
Psy-
chology,
1956,
52,
224-230.
GARNER,
W. R.
Uncertainty
and
structure
as
psychological
concepts.
New
York: Wiley, 1962.
GREEN,
D.
M.,
&
BIRDSALL,
T. G. The
effect
of
vocabulary size
on
articulation
score.
University
of
Michigan:
Electronic
Defence
Group,
Technical
Memorandum
No. 81 and
Technical
Note
AFCRC-
TR-57-58,
1958. (Reprinted
in J. A.
Swets,
Ed.,
Signal
detection
and
recognition
by
human
observers.
New
York:
Wiley,
1964.)
GREEN,
D.
M.,
&
SWETS,
J. A.
Signal
detection
theory
and
psychophysics.
New
York:
Wiley,
1966.
LUCE,
D.
Individual
choice
behavior.
New
York:
Wiley,
1959.
MELTON,
A.
Implications
of
short-term memory
for
a
general theory
of
memory. Journal
of
178
JOHN
MORTON
Verbal
Learning
and
Verbal
Behavior, 1963,
2,
1-21.
MILLER,
G.
A.,
HEISE,
G.
A.,
&
LIGHTEN,
W. The
intelligibility
of
speech
as a
function
of the
context
of
Ihe
test
materials. Journal
of
Experimental
Psychology,
1951,
41,
329-335.
MILLER,
N. E., &
BOLLARD,
J. C.
Social
learning
and
imitation.
New
Haven: Yale University
Press,
1941.
MORTON,
J. The
effects
of
context
on the
visual
duration threshold
for
words. British Journal
of
Psychology,
1964,
S5,
165-180.
(a)
MORTON,
J. The
effects
of
context upon speed
of
reading, eye-movements
and
eye-voice span.
Quarterly
Journal
of
Experimental
Psychology,
1964,
16,
340-354.
(b)
MORTON,
J. A
model
for
continuous language
be-
havior.
Language
and
Speech,
1964,
7,40-70.
(c)
MORTON,
J. A
preliminary
functional
model
for
language
behavior.
International Audiology,
1964,
3,
216-225. (Reprinted
in R. C.
Oldfield
&
J. C.
Marshall, Eds. Language. London:
Penguin
Books, 1968).
(d)
MORTON,
J. A
retest
of the
response bias explana-
tion
of the
word
frequency
threshold
effect.
Brit-
ish
Journal
of
Mathematical
and
Statistical Psy-
chology,
1968,
21,
21-33.
(a)
MORTON,
J.
Grammar
and
computation
in
lan-
guage
behavior.
Progress
Report
No. 6,
Center
for
Research
in
Language
and
Language Behavior,
University
of
Michigan,
May
1968.
(b)
MORTON,
J.,
&
BROADBENT,
D. E.
Passive
vs.
active
recognition models
or Is
your homunculus
really
necessary?
In W.
Wathen-Dunn
(Ed.),
Models
for the
perception
of
speech
and
visual
form.
Cambridge,
Mass.:
M. I. T.
Press,
1967.
NORMAN,
D. A.
Toward
a
theory
of
memory
and
attention.
Psychological
Review, 1968,
75,
522-
536.
POLLACK,
I.
Interaction between auditory
and
visual
information
sources
in
word identification.
Language
and
Speech,
1964,
7,
76-83.
POLLACK,
I.,
RUBENSTEIN,
H., &
DECKER,
L. In-
telligibility
of
known
and
unknown
message
sets.
Journal
of
the
Acoustical
Society
of
America, 1959,
31,
273-279.
POSTMAN,
L.,
&
ADIS-CASTRO,
G.
Psychophysical
methods
in the
study
of
word recognition.
Sci-
ence,
1957,
125,
193-194.
RUBENSTEIN,
H.,
&
POLLACK,
I.
Word predicta-
bility
and
intelligibility. Journal
of
Verbal
Learning
and
Verbal
Behavior,
1963,
2,
147-158.
STOWE,
A.
N.,
HARRIS,
W.
P.,
&
HAMPTON,
D. B.
Signal
and
context components
of
word-recogni-
tion behavior. Journal
of the
Acoustical
Society
of
America, 1963,
35,
639-644.
TREISMAN,
A.
M.,
&
GEFFIN,
G.
Selective atten-
tion:
Perception
or
response?
Quarterly
Journal
of
Experimental
Psychology,
1967,
19,
1-16.
TULVING,
E.,
MANDLER,
G.,
&
BAUMAL,
R.
Inter-
action
of two
sources
of
information
in
tachisto-
scopic
word recognition. Canadian Journal
of
Psychology,
1964,
18,
62-71.
TULVING,
E.,
&
PEARLSTONE,
Z.
Availability
versus accessibility
of
information
in
memory
for
words. Journal
of
Verbal
Learning
and
Verbal
Behavior,
1966,
5,
381-391.
WAUGH,
N.
C.,
&
NORMAN,
D. A.
Primary
memory.
Psychological
Review, 1965,
72,
89-104.
(Received January
19,
1968)
