Content uploaded by Douglas L Nelson
Author content
All content in this area was uploaded by Douglas L Nelson
Content may be subject to copyright.
How activation, entanglement, and searching a semantic
network contribute to event memory
Douglas L. Nelson &Kirsty Kitto &David Galea &
Cathy L. McEvoy &Peter D. Bruza
Published online: 4 May 2013
#Psychonomic Society, Inc. 2013
Abstract Free-association norms indicate that words are
organized into semantic/associative neighborhoods within
a larger network of words and links that bind the net togeth-
er. We present evidence indicating that memory for a recent
word event can depend on implicitly and simultaneously
activating related words in its neighborhood. Processing a
word during encoding primes its network representation as a
function of the density of the links in its neighborhood. Such
priming increases recall and recognition and can have long-
lasting effects when the word is processed in working mem-
ory. Evidence for this phenomenon is reviewed in extralist-
cuing, primed free-association, intralist-cuing, and single-
item recognition tasks. The findings also show that when a
related word is presented in order to cue the recall of a
studied word, the cue activates the target in an array of
related words that distract and reduce the probability of the
target’s selection. The activation of the semantic network
produces priming benefits during encoding, and search costs
during retrieval. In extralist cuing, recall is a negative func-
tion of cue-to-distractor strength, and a positive function of
neighborhood density, cue-to-target strength, and target-to-
cue strength. We show how these four measures derived
from the network can be combined and used to predict
memory performance. These measures play different roles
in different tasks, indicating that the contribution of the se-
mantic network varies with the context provided by the task.
Finally, we evaluate spreading-activation and quantum-like
entanglement explanations for the priming effects produced
by neighborhood density.
Keywords Activation .Quantum-like entanglement .
Semantic networks .Semantic memory .Working memory .
Priming .Extralist cuing .Word recognition .Reminding
Navigating a complex world requires many kinds of knowl-
edge, and we rely on fast, efficient access to information while
engaging hundreds of mental operations, including remem-
bering, thinking, reading, driving, and so on. The brain seems
to be wired to respond automatically to cues in its environ-
ment by simultaneously activating neighborhoods of related
knowledge. We see a friend and her name comes readily,
along with related information; we hear the word Atom,and
automatically electron,neutron,andbomb are activated. In
each instance, related word knowledge is activated, but most
of it remains in the shadow of awareness, available but not
apparent. Later, some cue reminds us of a prior event, and
activates it along with related knowledge; for instance, sup-
pose that we write a check to the electric company, seal it, and
put it on the desk. We get ready for work, rush out, and stop
the car at the end of the driveway after noticing the mailbox.
Our findings indicate that Check and Envelope remind us of
the associated words bill,money,mail, and the Mailbox cue
activates links to bill,mail,andmailman.Mailbox succeeds as
a cue because it implicitly activates words associated with a
momentarily forgotten task that reminds us to pay the bill.
In our research, we ask people to study words and then
we provide word cues to remind them of what they have
studied. We use free association norms to determine what
related words are likely to be activated when a word is
experienced, and in this article, we will describe how free
association can be used to measure the probabilities of links
between words, and how such links reveal a large and
D. L. Nelson (*)
Department of Psychology, University of South Florida,
Tampa, FL 33620, USA
e-mail: dnelson@usf.edu
K. Kitto :D. Galea :P. D. Bruza
Queensland University of Technology, Brisbane,
Queensland, Australia
C. L. McEvoy
School of Aging Studies, University of South Florida, Tampa, FL,
USA
Mem Cogn (2013) 41:797–819
DOI 10.3758/s13421-013-0312-y
organized associative/semantic network of words. Our mo-
tivating thesis is that the brain stores and retrieves experience
with objects and events, and words are invented to represent
such information. As experiences and the words that represent
them are repeated, they link together in interconnected seman-
tic neighborhoods. Such neighborhoods differ in both size and
connective density, with some neighborhoods being so dense
that they generate category names such as fruit,or“topics”
such as alcohol or drug abuse (Goldman, 1999;Griffiths,
Steyvers, & Tenenbaum, 2007; Stacy, 1997).
One purpose of this article is to show how the semantic
network contributes to episodic memory for recently experi-
enced word events. The findings show that experiencing a
familiar word during a study episode activates related words
that prime its activation state in this network and make it easier
to recall and recognize. We will review evidence from data
pooled over many experiments suggesting that cognitive pro-
cesses operate on this network, and we will evaluate equations
based on measures taken from the network in order to predict
recall in the presence of a related cue. We will show how the
effects of word knowledge change during encoding and re-
trieval, and with test delay and semantic context, and we will
discuss how priming effects related to automatic activation
can persist over long intervals. As a second purpose of the
article, we address the issue of how network activation works
when a word is processed during study. We will challenge the
spreading-activation explanation for priming effects and sug-
gest that experiencing a word simultaneously activates links
that encompass the word’s local neighborhood. Finally, we
will present a quantum-like entanglement model that de-
scribes and predicts how parallel activation captures neigh-
borhood density and the resulting priming effects.
The semantic network
Semantic/associative networks (shortened to semantic in this
article) can be constructed by presenting lists of words to sub-
jects asked to produce the first word to come to mind that is
meaningfully or associatively related (e.g., Cramer, 1968;
Deese, 1965; De Deyne & Storms, 2008a,2008b; Nelson,
Dyrdal, & Goodmon, 2005; Nelson, McEvoy, & Schreiber,
2004). Free association measures what one group of brains
knows about semantic relationships and predicts what other
groups of brains will be likely to know. It is a direct way to
measure what related words are likely to come to mind when a
word is experienced, and the validity of such measures depends
on how well they predict performance in word-processing tasks.
Free association to a word produces an array of associated
words that vary in probability and number. First-response as-
sociates are highly reliable (r= .89), even for responses pro-
duced by only two subjects, and we assume that they index a
word’s nearest neighbors (Nelson, McEvoy, & Dennis, 2000).
We collected such norms on 5,000 words over more than
20 years using more than 6,000 subjects, with an average of
about 150 subjects responding to each word. The words were
normed for addressing specific research questions that evolved
over time, and initially, we collected norms to study the effects
of cue–target relationships in the extralist-cuing task. In this
task, subjects study a list of unrelated words that are not directly
associated to one another, called targets (T), and the recall of
each target is prompted during testing by a semantically related
word cue (Q). Early interest focused on the effects of two free-
association measures, forward cue-to-target strength and back-
ward target-to-cue strength (see Table 1for abbreviations and
calculations of these and other variables). The experimental
findings showed that extralist cued recall increases as a positive
function of each link (e.g., Nelson, Fisher, & Akirmak, 2007):
Recall is more likely when a cue strongly activates a target in
free association, and in addition, an associate that is strongly
activated by the target is a more effective cue during test than is
a weakly activated associate.
Other findings showed that words vary in terms of how
many different associates they produce in free association, and
that the number of response alternatives produced by a word
that is used as a test cue has a robust but negative effect on
recall (called the cue set size effect in our experiments): The
more associates a cue has, the less the chance of recovering one
of its associates that has been studied as the target. This finding
generalizes across word types, taxonomic category names,
word stems, rhymes, and pictures (Nelson & Castano, 1984;
Nelson & McEvoy, 1979; Nelson, Walling, & McEvoy, 1979).
Cue set size, however, is an inclusive measure that includes the
target, indirect facilitating links that join the cue and target
(mediators and shared associates), and alternative responses.
For this article, we created a relative index that separates the
alternative responses responsible for negative effects from the
target and indirect links in statistical analyses of the data from
the earlier experiments. Alternative responses are now called
distractors, and as is indicated in Table 1, they are computed as
proportions that measure relative distractor strength; the prob-
ability of selecting a distractor instead of the target increases as
distractor strength increases.
Cue set size findings created a need to norm more words,
and this task became routine after learning that studied targets
with more links among their associates are more likely to be
recalled and recognized (e.g., Nelson, Bennett, Gee,
Schreiber, & McKinney, 1993; Nelson, McKinney, Gee, &
Janczura 1998; Nelson, Zhang, & McKinney, 2001). This
phenomenon was called connectivity in these experiments,
but to be more consistent with the literature on related topics,
we will refer to it in Fig. 1and elsewhere as neighborhood
density. Measuring neighborhood density required us to norm
a word in order to identify its closest associates, and then to
norm each of its associates using independent samples of
subjects. Figure 1a shows a toy network for word Tthat
798 Mem Cogn (2013) 41:797–819
produces three directly related associates, along with the four
links that connect them. Figure 1b illustrates this network as
an n×nasymmetrical adjacency matrix for word Tthat
represents its neighborhood. Placing T’s closest associates (A)
on the columns and again on the rows defines its neighborhood
in matrix form. Each of T’s associates is normed separately, and
we measure whether or not {1, 0} a link connects these
associates—for instance, reading along the rows, A
12
= 1 and
A
13
= 1, indicating that A
1
is linked to both A
2
and A
3
. Figure 1c
calculates neighborhood density by summing the number of
associate-to-associate links in the matrix and dividing by the
number of possible links. As with cue-to-target, target-to-cue,
and distractor strengths, neighborhood density is a relative
index of strength that varies from 0.0 to 1.0. Density increases
as the number of links among the target’s associates increase
relative to the size of its neighborhood. Neighborhood density
theoretically measures primed target strength, on the assump-
tion that targets with higher neighborhood densities are primed
to higher activation levels within the semantic network. For
example, Trombone has 85 out of 240 possible links among its
associates, whereas Basket has 10 out of a possible 342 links.
Their respective primed strengths are .354 and .029, and the
findings show that, with other variables controlled, words such
as Trombone are more likely to be recalled and recognized than
are words like Basket. In the last section of this article, we will
focus on neighborhood density in predicting recall and
Table 1 Variables shown to affect extralist cued recall, along with their computations
Variables Computation
Cue-to-target strength, S
qt
# subjects producing target in free association ÷ sample size
Target-to-cue strength, S
tq
# subjects producing cue in free association ÷ sample size
Cue-to-distractor strength, S
qd
# cue distractors ÷ # cue associates
*
Primed target strength, S
t’
# links among target’s associates ÷ # possible links
Mediator strength, S
m
# mediators ÷ # cue associates
Shared-associate strength, S
sa
# shared associates ÷ # target associates
Associate-to-target strength, S
at
# associate-to-target links ÷ # target associates
Target-to-distractor strength, S
td
# target distractors ÷ # target associates
*
Distractors exclude the target, mediators, and shared associates. Mediators provide indirect links from the test cue to the target (e.g., Proton →
electron →Atom). Shared associates occur when the cue and target have the same associate in common (e.g., Proton →electron and Atom →
electron).
a. Graph depicting Word T, its 3 associates, and their 4 connecting links
b. Asymmetric adjacency matrix depicting links among T’s associates
Normed Associates of T
Target T
A1
A2
A3
Associates
A1
---
1
1
Normed
A2
1
---
1
As Cues
A3
0
0
---
c. Primed target strength, St’
A2
Target T A1
A3
ΣΣ
Fig. 1 Measuring primed target
strength on the basis of neigh-
borhood density
Mem Cogn (2013) 41:797–819 799
recognition and propose a quantum-like model for explaining
its effects.
What does free association measure?
In free association, response production is not really “free,”in
the sense that any response is always acceptable. Free associ-
ation is restricted because the instructions bias responses, from
the very general—any word—to the very specific—category
instances, properties, rhymes, feelings, topics like “alcohol
makes me feel,”and so on (e.g., Battig & Montague, 1969;
Cramer, 1968; MacRae, Seidenberg, & McNorgen, 2005;
Reich & Goldman, 2005). Because the task is so familiar,
individuals can access and generate any type of related word
on the basis of instructions that set a comprehensible goal that
defines a context for eligible words. Different biases produce
different semantic maps, and at this time, no single “best”map
predicts the effects of all types of mental operations on all
types of word knowledge (Maki & Buchanan, 2008). For our
norms, we requested a response that reflects meaning or
association, which allowed responses to be based on a variety
of semantic relationships.
The second word in the task name, association,isalso
misleading, because it implies that cue-to-target strength is
determined only by an association formed by coexperiencing
words Qand Tin the same context. Association by learning is
critical, but observing the existence of a link reveals nothing
about its source. Free association indicates the probability that
one word will lead to production of another and tells us
nothing about the information used to produce the response.
The observation could be generated through learned associa-
tion, recency, imagery, semantic similarity, or contrast.
Several studies have shown that semantic information
concerning categories, properties, emotions, and so on, is well
represented in free-association data (Borge-Holthoefer &
Arenas, 2010; Brainerd, Yang, Reyna, Howe, & Mills, 2008;
De Deyne & Storms, 2008a; Hutchison, 2003;Monaco,
Abbott, & Kahana, 2008; Steyvers, Shiffrin, & Nelson,
2005). Most importantly, when we refer to the “semantic
network,”we are referring to the normalized network that
emerges from our free-association procedure. Furthermore,
we assume that every person has a variant of such a network
in memory and that processes such as activation and search
operate directly on this representation.
Organization of the semantic network
In many sciences, networks tend to be large, asymmetric, and
irregular structures that are well suited for describing complex
systems (Newman, 2003). Networks describe links between
people (Moreno, 1934), computers (Barabási, 2002), neurons
(Watts & Strogatz, 1998), metabolic states (Jeong, Tombor,
Albert, Oltvai, & Barabási, 2000), and words (Steyvers &
Tenenbaum, 2005). Their topological structure helps scientists
understand how different kinds of entities communicate in
interactive systems, and they have been used to describe a
wide range of processes, including percolation, spread of
disease, phase transitions, protein interactions, and computer
search, as well as activation and search (Barabási, 2002;
Nelson et al., 1998;Newman,2003). Networks are also useful
to scholars investigating complexity itself (e.g., Barabási,
2002; Palla, Derényi, Farkas, & Vicsek, 2005; Steyvers &
Tenenbaum, 2005; Strogatz, 2003; Watts & Strogatz, 1998).
Research across disciplines has indicated that different
networks are organized in strikingly similar ways. By treating
networks as graphs and borrowing the terminology and math-
ematics of graph theory, researchers have shown that networks
built from different entities and types of links tend to have
similar statistical profiles. Of particular interest, Steyvers and
Tenenbaum (2005) used our norms to create a semantic net-
work by generating a 5,018 × 5,018 word matrix in which
normed cues were represented on the rows and responses were
shown on the columns, as in Fig. 1b. Their analysis shows that
the network is sparsely linked, with each word being directly
connected to an average of only 12.7 other words.
Furthermore, short path links are characteristic of all small-
world networks (Watts & Strogatz, 1998). The average direct-
ed path length from any one word in our norms to any other
takes only 4.27 links or “steps,”so the semantic network
qualifies as a small-world network. More importantly for this
article, a property termed clustering in graph theory is
described in this article by neighborhood density, the relative
number of links among a word’sassociates(Fig.1c).
The network analyses indicated that words in the semantic
network are directly connected to only a few other words, but
on average, they are connected to nearly all other words in the
net within a few mediating links. The network has a small-
world structure composed of words with varying levels of
neighborhood density that are linked by short paths to and
from other words. As a network, words are represented as a
system of nodes and links, as opposed to isolated pairs, and
theoretically, a mental process operating on one word in the
system has the potential to change the states of related words
in the network. The system in which words reside is distrib-
uted, but many alternative paths can lead to any single word,
and short paths provide shortcuts that may affect the activation
states of distant concepts (Watts & Strogatz, 1998).
We assume that the semantic network is a dynamically
evolving, complex system. Complexity arises because we
know thousands of words that are linked in semantic mem-
ory in ways that can affect how well a given word will be
remembered. Dynamics arise because portions of the system
are always growing or weakening in response to experience
(Hills, Maouene, Riordan, & Smith, 2010;Steyvers&
Tennenbaum, 2005). The network is also dynamic in terms
of how it responds to the contexts provided by different
800 Mem Cogn (2013) 41:797–819
tasks. We next review findings showing that a target’s
neighborhood density has different effects, depending on
the task and the semantic context in which it is experienced.
Knowledge × Context interactions
In this section, we review how the semantic network is
engagedinavarietyoftasks, including cued recall,
primed free association, intralist cuing, and word recog-
nition. We initially focus on the extralist-cuing task,
resembling the mailbox example above, because of its
importance to everyday memory events and because it
is more interesting from a network perspective. Pooling
the results over dozens of extralist-cuing experiments
provides a database on thousands of word pairs that
we will use to identify the most important network
variables and how they can be incorporated into an
equation that effectively predicts recall.
Extralist cuing
During the study phase of the standard extralist task, a list
of unrelated familiar words is presented one at a time for 3 s,
and subjects read each word aloud and attempt to remember
as many as possible. Test instructions are provided immedi-
ately after the last study word. During testing, another set of
words is presented one at a time at a self-paced rate, and the
instructions ask subjects to use each cue to recall a related
target word from the study list (e.g., use the cue “Proton”to
recall the target “Atom”). The procedure is called extralist
because the test cues are physically unavailable during
study, and hence, successful recall depends on the cue
accessing the target in the semantic network. This charac-
teristic makes extralist cuing an ideal “reminding”paradigm
for studying how prior knowledge captured by a single word
can be used to respond to the environment.
Tab le 1presents a list of eight variables shown to affect
recall in extralist cuing that have been incorporated into a
statistical prediction algorithm called Pier2,standingfor“pro-
cessing implicit and explicit representations”(see, e.g.,
Nelson, Goodmon, & Ceo, 2007, and Nelson & Zhang,
2000, for description and computational examples). One pur-
pose of this article is to report a simplified computation that
grew out of two problems with Pier2: It overpredicts recall,
and the eight variables in its prediction equation are not on the
same scale, with some representing strengths, such as cue-to-
target strength, and others representing sums of strengths.
Although the equation is well correlated with recall, we
thought that transforming each variable to the same scale
would solve the over prediction problem. Just as importantly,
given correlations among the variables, not all of them may be
necessary for predicting recall thus simplifying the problem of
incorporating them into a prediction equation. As we will
show, transforming our past measures into proportions unex-
pectedly forced us to reconceptualize Pier and to revise our
interpretation of both cue and target set size effects.
All eight variables shown in Table 1were transformed to
free-association-determined strengths that ranged from 0.0 to
1.0 and were entered into an extralist-cuing database com-
posed of 4,068 cue–target pairs rescored at the item level.
1
In
the articles that reported the original experiments, the data
were pooled over items, and the usual subject-based analysis-
of-variance procedures were applied to determine what vari-
ables were significant. In contrast, in this article, the data are
pooled over subjects in all analyses. The number of subjects
who correctly recalled a target relative to the total number who
studied the target computes the probability of recalling each
cue–target pair (see Nelson & Zhang, 2000). Both subject and
item analyses indicate whether a variable is important, but
item analyses are critical for predicting the probability of
recalling a given cue–target pairing. This type of analysis
allowed us to determine the influence of a variable over
thousands of cue–target pairs, which is important when the
goal is to determine how and how much the network contrib-
utes to memory for recent word events. In addition, the item
analysis allowed us to use regression analyses that statistically
controlled for correlations among the variables.
In the database, the cue–target pairs were shown in two
columns, with other columns representing the observed and
predicted probabilities of correct recall, information on how
the list was studied, measures of the eight variables, and
other related information. The analyses were restricted to
2,803 pairs studied in the standard task, and a preliminary
analysis was used to determine whether the new scaling
procedure eliminated some variables when correlations with
other variables were statistically controlled in a simulta-
neous multiple regression. Probability of correct recall was
the dependent measure, and measures of the eight variables
served as predictors.
The preliminary analysis results indicated that the first four
variables shown in Table 2were more important for predictive
purposes. The regression was significant [F(8, 2794) =
205.30, MS
res
= .042, R= .61, R
2
= .37], and the partial
correlations shown in the bottom row of Table 2indicate that
the variables represented in the first four columns were more
strongly correlated with recall than were the last four. The four
variables having the largesteffects were cue-to-target strength,
S
qt
, target-to-cue strength, S
tq
, cue-to-distractor strength, S
qd
,
and neighborhood density as a measure of primed target
strength, S
t'
. Of the last four variables, shared associate
strength and associate-to-target strength failed to have signif-
icant effects in this analysis, and mediator strength and target–
distractor strength had reduced effects because of correlations
1
All databases mentioned in this article are available on request from
the first author.
Mem Cogn (2013) 41:797–819 801
with other measures. These four variables were dropped from
the prediction equations that will be described later, because
they were redundant and reduced predictability when includ-
ed. The failure of associate-to-target strength to affect recall is
important to spreading-activation theory, and we will return to
this result later. The failure of target–distractor strength to
achieve significance deserves special notice, because this
variable played a key role in our earliest investigations of the
semantic network. Many experiments had focused on the
effects of varying target set size because it has robust negative
effects on cued recall, but as with cue set size, it is correlated
with several variables that have positive effects (see Table 2).
However, we are not dismissing target set-size effects as
irrelevant. In the conversion to proportions, target set size
determined the size of the associative matrix, and hence the
relative magnitude of neighborhood density. Given a fixed
number of links among the target’s associates, neighborhood
density decreases as target set size increases, so the negative
effects of target set size are built into the density index.
The first row of Table 3shows that the results of a second
analysis, limited to the four most important predictors of
recall, are significant overall, and that each predictor has
significant effects. The correlations, partials, and descriptive
statistics are shown in Table 4. There, the partial correlations
are shown in the bottom row, and each variable is signifi-
cantly correlated with recall after correlations among the
variables are statistically controlled. Of the variance that
could be explained, a separate, unforced stepwise regression
showed a highest-to-lowest ordering of cue-to-target strength
S
qt
= 58.9 %, neighborhood density S
t'
= 20.8 %, cue-to-
distractor strength S
qd
= 11.8 %, and target-to-cue strength
S
tq
= 8.5 %. Recalling the target depends most on the cue-to-
target strength S
qt
, and next-most on neighborhood density
S
t'
. Target-to-cue strength S
qt
has a smaller positive effect, and
Table 2 Correlations (above diagonal) and partial correlations (below diagonal) among the eight network variables that affect extralist cued recall
S
qt
S
tq
S
qd
S
t’
S
m
S
sa
S
at
S
td
rWith Prob. Recall
S
qt
—–.01 –.29 –.04 .27 .04 .17 –.02 .46
S
tq
.01 —.11 .20 –.20 .18 –.21 –.30 .19
S
qd
–.13 .13 —–.16 –.73 –.43 –.11 .37 –.37
S
t’
–.10 .08 .07 —.08 .63 .10 –.58 .26
S
m
–.02 .04 –.70 –.07 —.23 .45 –.15 .30
S
sa
.00 –.13 –.15 .34 .02 —–.04 –.87 .34
S
at
.11 –.14 .31 .22 .51 –.02 —.11 .10
S
td
.06 –.25 .08 –.06 .05 –.76 .08 —–.33
Partial rWith Prob. Recall .42 .19 –.10 .11 .06 .02 .02 –.07
Table 3 Results of regression analyses evaluating the effects of the four predictors as a function of memory task
FMS
res
RAdj R
2
S
qt
S
tq
S
qd
S
t’
Study Cond.
Extralist Cuing F(4, 2803) = 387.36 .043 .597 .355
Std. Coeff. .395 .183 –.246 .194 –
SE .034 .028 .022 .034 –
tValue 24.81 11.64 15.08 12.26 –
Primed Free Assoc. F(5, 464) = 47.42 .019 .583 .332
Std. Coeff. .441 .162 –.225 .081 .145
SE .100 .069 .037 .087 .007
tValue 10.77 4.19 –5.18 2.07 3.76
Intralist Cuing F(4, 509) = 13.17 .037 .306 .087
Std. Coeff. .263 .011 –.090 –.056 –
SE .085 .052 .049 .001 –
tValue 5.78 .246
†
–1.91
†
–1.24
†
–
Recognition F(3, 1647) = 221.02 .919 .536 .286
Std. Coeff. –– – .071 .381
SE –– – .143 .033
tValue –– – 3.37 18.33
†
Not significant
802 Mem Cogn (2013) 41:797–819
only cue-to-distractor strength S
qd
has negative effects on
recall: All things being equal, a cue that activates a higher
proportion of distractors is more likely to fail.
Predicting extralist cued recall
Retrieving a target event from the semantic network can be
direct when cue-to-target strength is very strong, but retriev-
al becomes a problem as the cues become weaker and
retrieval less certain. For the most part, in our experiments
cue-to-target strength was generally weak (averaging .15)
because we were manipulating several variables in factorial
designs and needed to keep recall below ceiling. Given the
uncertainty of retrieval, a test cue was more likely to utilize
all of the available information. In the Pier2 framework,
measures of variables known to affect recall had been com-
bined in a single prediction equation. We updated this equa-
tion now because of the changes in the number of key
measures and in their underlying scales, and we will refer
to the new approach as Pier3. There are many ways to
combine four predictors, and we evaluated many plausible
computations. The equation that we report below is the best
general predictor, and although it does not constitute a
model in the same sense as general memory models (e.g.,
the search of associative memory [SAM] model), it carries
implications for how a cue reminds us to recover a recent
word event from the semantic network.
Figure 2illustrates the positive and negative effects in-
volved when a test cue activates its associates, given the
goal of recalling a related and recent word event. We assume
that the four sources of strength represented in the figure act
together in affecting probability of recall. The target activates
its associative network, and this activation strengthens the
target’s representation in semantic memory. By activating an
array of related associates, the test cue narrows the range of
potential targets, and the target is selected from within this
array. Cue-to-target (S
qt
) and target-to-cue (S
tq
) strengths lo-
cate the primed target in the array, and its recall is more likely
when these links are stronger and when the target has been
primed to a higher level of activation during study (S
t'
).
Theoretically, target priming makes its representation more
distinct within the array of associates activated by any related
cue and increases the likelihood of selecting the target. In
contrast, distractors have negative effects on recall, and cue-
to-distractor (S
qd
) strength is interpreted as sampling error. As
distractor strength increases, the probability of sampling a
distractor as opposed to the target increases.
Equation 1brings the four sources of information about the
target together. Words in the semantic network exist in a tangle
of correlated links loosely organized into a complex system,
and recall in extralist cuing is based on more than just how
strongly the test cue is related to its target. Equation 1is a
SAM-like equation stating that the semantic network’scontri-
bution to target recall in extralist cuing can be predicted by
computing the probability of recalling a primed target, given
the test cue. P(T′/Q) is a ratio of positive sources of
Table 4 The four strongest predictors of extralist cuing, with correlations (above diagonal) and partial correlations (below) in the first five
columns, and descriptive statistics in the last three columns
S
qt
S
tq
S
qd
S
t’
Prob. Recall Mean SD Range
S
qt
––.01 –.29 –.04 .46 .152 .120 .00–.818
S
tq
–.05 –.11 .20 .19 .099 .144 .00–.917
S
qd
–.15 .21 ––.16 –.37 .605 .192 .091–1.00
S
t’
–.18 .17 –.13 –.26 .152 .118 .00–1.00
Prob. Recall .43 .22 –.27 .23 –.518 .258 .00–1.00
Test Cue Q
Distractor
Distractor
Primed
Target St’
Distractor
Distractor
Distractor
Stq
Sqt
Fig. 2 Recovering primed target T' from its test cue Q. Cue-to-target
strength (S
qt
) is represented in the arrow from the test cue to the target,
and target-to-cue strength (S
tq
) is represented in the arrow from the
target to the test cue. Both strengths are taken directly from free-
association norms. Primed target strength (S
t'
) is calculated, as an
example, in Fig. 1(see also Eq. 4in the text). The cue–distractor
strength in this figure is determined by dividing the number of
distractors by the total number of responses produced to the cue in
free association, after eliminating mediators and shared associates.
Mediators and shared associates facilitate recall and are not counted
as distractors, and because there are none in the figure, the cue–
distractor strength is 1/6 or .167
Mem Cogn (2013) 41:797–819 803
information about the target, relative to the total amount of
information activated (see, e.g., Raaijmakers & Shiffrin,
1981). The probability of recalling a primed target is a func-
tion of total cue–target strength, relative to the total amount of
information activated by the test cue, which includes total
cue–target strength and distractor strength. Total cue–target
strength S(Q,T′)iscapturedinEq.2by summing the cue-to-
target and target-to-cue strengths and adding the result to the
target’s neighborhood density. Cue–distractor strength is de-
fined in Eq. 3as the number of distractors relative to the
number of associates activated by the test cue (cue set size).
Finally, Eq. 4measures target priming as a function of its
relative neighborhood density.
All of these sources of information are embedded in the
semantic network, and we assume that target priming and
cue-initiated search operate directly on this network. The
predicted probability of recalling a primed target from the
semantic network, given its test cue, is
PT
0=QðÞ¼ SQ;T0
ðÞ
SQ;T0
ðÞþSQ;DðÞ
;ð1Þ
where total cue–target strength S(Q,T')is
SQ;T0
¼Sqt þStq þSt0;ð2Þ
cue–distractor strength S(Q,D)is
SQ;DðÞ¼
Pd
i¼1Di
m;ð3Þ
where d= distractors and m= number of cue associates, and
target-priming probability P(T') for associates A
ij
is
St0¼Pn
i¼1Pn
j¼1Aij
nn1ðÞ
;ð4Þ
where n= number of target associates.
Results
To test the predictions of Eq. 1, we used the extralist-cuing
database for the cue–target pairs learned under standard condi-
tions (i.e., 3-s study, “remember”instructions). The probabili-
ties of observed and predicted recall were normally distributed
around means of .518 (SD = .26) and .392 (SD = .16), respec-
tively. We computed the error distribution by subtracting pre-
dicted from observed recall and plotting the results, and this
distribution was normal and shifted to the right because the
equation underestimates recall.
The first row of Table 5presents the results of a multiple
regression analysis with Eq. 1as the sole predictor, and it is well
correlated with probability of recall, r= .567. Figure 3shows
observed and predicted recall as a function of four categories
according to Eq. 1, and as can be seen, it underestimates the
observed recall by about the same amount in each category. The
transformation to proportions solves the overprediction prob-
lem associated with Pier2. Equation 1uses fewer variables,
predicts as well as Pier2, and explains 32.1 % of the total
variance in extralist cuing. When the correlation is corrected
for the reliability of the free-association and extralist-cuing
tasks (Myers & Well, 1995, pp. 394–395, 472–473), the corre-
lation between predicted and observed recall rises from r=.567
to r= .703, and the explained variance rises to 49.5 %
(see Nelson & Zhang, 2000,p.610).Thus,nearlyhalfofthe
total variance of extralist recall is explained by free-association
measures based on links among the words comprising the
semantic network. The network can contribute substantially to
episodic recall.
Effects of test delay
Fifteen experiments showed that recall in the extralist-cuing
task is negatively affected by test delays filled with solving
multiplication problems (e.g., Nelson, Bajo, & Casanueva,
Table 5 Results of regression analyses for Eq. 1as the sole predictor, in the first row, and regression comparisons between Eq. 1against test delay,
presentation rate, and study instructions in the successive rows
FMS
res
RAdj R
2
Std. Coeff. SE t Value
Regression F(1, 2801) = 1,326.00 .045 .567 .321
Eq. 1.567 .025 36.41
Regression F(2, 2552) = 285.07 .048 .427 .182
Eq. 1.384 .026 21.43
Test delay –.164 .000 –9.15
Regression F(2, 3220) = 778.08 .045 .571 .325
Eq. 1.565 .023 39.02
Rate .084 .010 5.81
Regression F(2, 369) = 930.25 .045 .597 .356
Eq. 1.551 .022 39.78
Instructions .207 .009 14.93
804 Mem Cogn (2013) 41:797–819
1985; Nelson et al. 2007;Nelsonetal.,1998). Their data were
rescored at the item level and pooled over experiments in
order to portray the effects of delays at 2, 7, 10, 12, 22, and
1,442 min. Two minutes were added to each delay in order to
accommodate study-to-test time that would include time for
reading the test instructions. A multiple regression analysis
that included the four variables plus test delay as predictors
was significant [F(5, 2549) = 138.82, MS
res
=.046,R=.46,
R
2
= 21.2 %]. The coefficients or weights allocated to cue–
target strength S
qt
, target–cue strength S
tq
,cue–distractor
strength S
qd
, neighborhood density S
t'
, and delay by the re-
gression equation were all significant, and the respective
correlations between each source and observed recall were
.32, .14, –.21, .22, and –.19. When the interactions between
each source and delay were included as predictors, none were
significant. The absence of delay interactions is inconsistent
with experimental findings showing that the strong–weak
difference for each variable declines as delay increases
(see also Sahakyan & Goodmon, 2010). Examinations of
mean recall with the data split along strong–weak levels for
each variable indicated that this trend was apparent in the
database, but the effects, though smaller, were apparent even
after 24 h. Finally, as is shown in the second row of Table 5,
Eq. 1and delay were significant predictors in a simultaneous
multiple regression, and each predictor was significantly cor-
related with recall, rs=.40and–.19.
Alog–log plot of probability of recall and test delay
approximated a straight line, and a power function fitted to
recall as a function of untransformed values explained 88 %
of the variance that could be explained, with an intercept
constant of .541 and a slope of –.109. Effects of delay were
incorporated into Eq. 1by weighting it by t
–αc
, where tis
time in minutes, αis decay rate estimated by the power
function, and cis a context parameter (c≥1) that is esti-
mated from the data and accounts for differences in how the
retention interval is filled (e.g., with multiplication prob-
lems, studying additional lists, or room and experimenter
changes). When the effects of test delay are taken into
account, Eq. 1becomes
PT
0=QðÞ¼ SQ;T0
ðÞ
SQ;T0
ðÞþSQ;DðÞ
t/c
ðÞ:ð5Þ
Figure 4shows probability of recall as a function of test
delay using Eq. 5with c= 1. Recall declines sharply over
the first 20 min and levels out after 24 h. When Eq. 5was
used as a single predictor of recall in the multiplication
problem database, it emerged as a significant predictor,
F(1, 2553) = 686.5, MS
res
= .046, R= .46, R
2
= 21.2 %.
We attribute the effects of test delay to decay and the loss of
context information (e.g., Nelson & Goodmon, 2003;Nelson,
Goodmon, & Akirmak, 2007;Nelsonetal.2007;Sahakyan&
Goodmon, 2010). Theoretically, the target is primed and
encoded in conjunction with environmental context cues sam-
pled during study, and asking subjects to switch attention from
the study list to math problems and the like disrupts the
recovery of these cues. Additional support for the importance
of context has come from studies showing that changes in the
physical environment—produced by changing rooms, exper-
imenters, or both between study and test—as well as requiring
the study of additional lists during the delay, reduce the
probability of recall. However, neighborhood density effects
are attributed to target priming. Forgetting describes why
recall declines, but it does not explain why priming effects
are reduced but still apparent after 24 h, when activation
normally decays within milliseconds in speeded lexical and
semantic decision tasks. In contrast to the case in speeded
tasks, density-induced target-priming effects somehow persist
in extralist cuing, and we suggest a reason below.
Effects of study condition
Study conditions varied by means of presentation rate and
instructions have robust effects on recall in extralist cuing. The
database included 1,265 pairs encoded under nonstandard
conditions. In some experiments, presentation rate was varied
under the otherwise standard conditions, and recall increased
with slower rates, averaging .47 (SE =.02,n= 210 pairs), .52
(SE =.01,n= 2,803), and .60 (SE =.02,n= 210) for rates of
1.5, 3.0, and 6.0 s (see, e.g., Nelson, Schreiber, & McEvoy,
1992). With Eq. 1and rate as predictors, the results in the third
row of Table 5indicate that the regression and both predictors
are significant. The correlations with recall for Eq. 1and rate
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
.00-.20 .20-.40 .40-.60 .60-.80
Probability of Recall
Predicted Recall in Four Categories
OBSERVED RECALL PREDICTEDRECALL
302 1,253 898 301
Fig. 3 Observed and predicted probabilities of recall as a function of
the Eq. 1predictions, in four categories
Mem Cogn (2013) 41:797–819 805
are r= .57 and .08, and they explain 98.1 % and 1.9 % of the
variance that can be explained. In extralist cuing, the contri-
bution of semantic memory is substantially greater than the
contribution of additional study time.
In other experiments, different groups of subjects have
processed the targets by naming their vowels, trying to re-
member them in the standard condition, or rating their con-
creteness (e.g., Nelson et al. 2007). Recall increased with
elaboration and averaged .39 (SE =.02,n= 210), .52
(SE =.01,n= 2,803), and .68 (SE =.02,n= 427) for naming
vowels, “remember”instructions, and rating concreteness. We
coded instructions as –1, 0, and +1 and entered it along with
Eq. 1as a predictor in a multiple regression. Table 5shows
that the regression equation was significant and that Eq. 1and
the study instructions contributed significantly to recall.
Equation 1and instructions were correlated with recall at
r= .56 and .23, and an unforced stepwise regression indicated
that they explained 87.9 % and 12.1 % of the variance that
could be explained.
Consistent with study time, preexisting information in the
semantic network contributes more to recall in extralist
cuing than does level of processing. Recall in extralist cuing
relies on both semantic and episodic manipulations, but it
depends more on the contribution from the semantic net-
work than on explicitly defined encoding procedures. The
network captured in Eq. 1explains more variance than other
predictors because the test cue is absent during study and
target recall hinges on using the test cue to access the target
within the semantic network. In experiments, these two
sources of information generally do not interact, but when
they do, the interaction is weak and potentially the result of
ceiling effects. These results suggest that semantic network
variables and episodic encoding operations (E) have addi-
tive effects on recall in the extralist task, and that total cue–
target strength (Eq. 2) can be re-represented as
SQ;T0
ðÞ¼Sqt þStq þSt0þE:ð6Þ
Developing an equation for Eis beyond the scope of this
article, but we will consider two implications of Eq. 6. The
first is that encoding operations applied according to the
study instructions add their effects directly to the total cue–
target strength S(Q,T′).Total strength is greater when the
study instructions focus on semantic processing or remem-
bering rather than on naming vowels. The second implica-
tion is that two different types of information contribute to
target recall in the extralist-cuing task, one based on episod-
ic encoding operations and one based on implicitly activated
semantic information. The episodic processes are tied to
how targets are encoded, to study time, and to general
context cues, and hence to recollective retrieval processes.
The implicit processes are linked to the strength of the cue–
target relationship and to target priming, and hence to auto-
matic, unaware retrieval processes.
Pier incorporates both dual encoding and dual retrieval
processes in explaining extralist cuing. Pier3 assumes that
the appearance of a target can prime its representation in the
semantic network in accordance with its neighborhood
Fig. 4 Observed and predicted
probabilities of recall as a
function of minutes of test delay
806 Mem Cogn (2013) 41:797–819
density, and that regardless of its primed state, a target is
automatically loaded into working memory and subjected to
encoding operations controlled by the study instructions.
Both target-priming effects and the encoding processes ap-
plied to the target in working memory strengthen its repre-
sentation in the semantic network. As is illustrated in Fig. 5,
the studied word Atom is represented in the network, and its
appearance in the experimental setting activates this repre-
sentation, with neighborhood density theoretically affecting
its primed strength. The application of working memory
operations to a primed representation of Atom adds to its
strength in the semantic network, so that any cue linked to it
in the network has an increased likelihood of selecting it as a
response during testing. Importantly, we assume that Atom’s
associates serve only to prime its representation in semantic
memory, and that they are not processed in working mem-
ory. We found small implicit associative strengthening ef-
fects in an early experiment (Nelson, Bajo, McEvoy, &
Schreiber, 1989), but several unpublished experiments by
Schreiber failed to replicate such effects.
As is suggested in Fig. 5, primed targets such as Atom are
encoded in the presence of elements of the environmental
context that are simultaneously represented in working
memory, so context becomes associated with the primed target
(cf., Raaijmakers & Shiffrin, 1981). Atom’s representation in
the network is temporarily strengthened by all sources of
information, including priming, explicit encoding operations,
and context. When Proton is presented as a reminding cue, it
initiates a search via its associative array. Atom is now more
likely to be selected than are Proton’s other associates, if it has
been primed and explicitly encoded via study instructions, and
this encoding occurs in the presence of information about the
context of the experience. Relative to Proton’s distractors,
Atom now has a greater chance of being recalled, because its
state in the semantic network has been strengthened via both
episodic and implicit processes, and this state explains why
target-priming effects can be long-lasting in extralist cuing:
They last when target words are encoded in a primed state in
working memory.
Primed free association
Primed free-association and extralist-cuing tasks are similar
because the same list of targets and test cues can be used
along with the same study conditions, and the hypothesis
that primed information is retrieved from semantic memory
Fig. 5 The representation of
Atom in semantic memory can
be strengthened by priming
effects produced by its
neighborhood density in the
network and by encoding
processes in working memory.
Although priming effects are
generally short-lived, they can
be long-lasting when a primed
version of a target word is
encoded in working memory
Mem Cogn (2013) 41:797–819 807
predicts that the same variables will be important in both
tasks. The main difference between these tasks occurs dur-
ing testing. In primed free association, subjects are asked to
free-associate the first related word to come to mind for each
test cue, under a cover story suggesting that norms are being
collected for another experiment, and they are not told that
the test cues are related to the targets. To reduce the prob-
ability of their discovering the relationship, the test often
begins with unrelated “cues”andthencontinueswitha
random mix of equal numbers of related and unrelated cues
(e.g., Chappell & Humphreys, 1994; Humphreys, Tehan,
O’Shea, & Bolland, 2000; Nelson & Goodmon, 2002;
Zeelenberg, Shiffrin, & Raaijmakers, 1999). We focus on
the probability of recalling the target as a function of the
variables that comprise Eq. 1, to determine whether they
have similar effects in extralist cuing and primed free asso-
ciation. The intentional-recall instructions used in extralist
cuing make subjects aware of the relevance of the target
words just seen in the study context and the cue–target
relationship, whereas free-association instructions relax
awareness on both counts. Neither the study context nor
the semantic relationships between the cues and targets are
ever mentioned in the test instructions. Relaxing awareness
decreases recall, but the variables that affect intentional
recall should have similar effects in primed free association
if both tasks are relying on the same semantic network.
Results
Previous experiments confirmed this expectation and showed
that target recall was greater when S
qt
and S
tq
were stronger, as
well as when S
t'
was higher. As in the extralist-cuing task, they
also showed that target recovery is more likely when attention
is directly focused on target meaning during study as com-
pared to standard “remember”instructions, and when testing
is immediate as opposed to delayed (Goodmon & Nelson,
2004; Nelson & Goodmon, 2002;Nelsonetal.2007).
The data from these and related experiments involving
immediate tests were rescored at the item level and used to
create a primed free-association database that included 470
cue–target pairs, probability of target recall, and the predictor
measures. Although recall is very low in primed free
association, the results of a simultaneous multiple regression
analysis, shown in Table 3, indicated that the regression was
significant, as were the effects of each of the four predictors.
Tab le 6presents the correlations for each variable, and note
that although S
qd
was not varied in any experiments, it none-
theless emerged as a significant negatively correlated predic-
tor. As in extralist cuing, target recallincreased significantly as
a positive function of S
qt
,S
tq
,andS
t'
, and it declined as cue-to-
distractor strength S
qd
increased. An unforced stepwise regres-
sion showed that recall was significantly affected by all four
network variables, with S
qt
,S
tq
,S
qd
,andS
t'
explaining 23.9 %,
4.0 %, 3.2 %, and 0.4 % of the total variance. The effects of all
four variables were reduced in primed free association as
compared to extralist cuing, but each variable was significant,
and the effects of S
t'
, although quite small, indicate that target
priming still affected recall.
Predicting priming
A multiple regression analysis with probability of target recall
as the dependent measure and Eq. 1as the predictor was
significant, but it overpredicted recall, F(1, 468) = 118.36,
MS
res
=.023,R=.45,R
2
= 20 %; the mean predicted and
observed recall were .33 and .19. Observed recall is very low
in primed free association when directly compared to extralist
cuing, where recall can be twice as high (Nelson et al. 2007).
This low recall arises because subjects are unaware of the
importance of the study episode and the cue–target relation-
ships. Weighting Eq. 1by .5 brings the predicted mean target
recall more in line with the observed recall, and although it
must be adjusted, Eq. 1explains 20 % of the total variance and
is an acceptable predictor of primed free association. Target
recall is thus based on the semantic network, even in the
absence of an intent to recall.
Studying semantically related pairs
In extralist cuing, each target is presented in the absence of
semantic context during study, but in this section we will
consider a task in which the target is studied in the context
of a semantically related word. A word seen in either context
implicitly activates its closest associates (e.g., in lexical
Table 6 Correlations (above diagonal) and partial correlations (below) for primed free association
S
qt
S
tq
S
qd
S
t’
Study Condition rWith Prob. of Recovery
S
qt
––.10 –.37 .02 –.13 .49
S
tq
–.12 –.12 .07 .12 .11
S
qd
–.21 .15 ––.31 .17 –.37
S
t’
–.14 .10 –.29 ––.11 .16
Study Condition –.14 .07 .13 –.09 –.06
Partial rWith Prob. of Recovery .45 .19 –.23 .09 .17 –
808 Mem Cogn (2013) 41:797–819
decisions, both meanings of a polysemous word are mo-
mentarily activated; Kintsch, 1988). However, the effects of
this broad activation on recall could not be more different in
the two contexts. When seen in isolation in the extralist task,
target meaning is uncertain, and the activation of related
words provides an internally generated semantic context that
primes the target’s representation to varying degrees.
Theoretically, this primed representation is loaded into and
studied in working memory, thus preventing the rapid decay
of target priming. For example, when Bolt is studied in
isolation, both its “fastener”and “storm”meanings support
and prime its activation state, and because the target is
rehearsed and encoded in this state, both meanings affect
recall, even when Bolt is cued by Screw or Lightning, which
specify one meaning or the other (Gee, 1997). Both mean-
ings contribute to Bolt’s priming level during study, so they
affect recall even when search is initiated by only one of its
meanings.
In contrast to semantic isolation, encountering a pair of
semantically related words during study tends to produce a
semantically specific encoding (Tulving & Thomson, 1973).
For example, in the intralist-cuing task, after studying
Screw–Bolt,Screw is an effective cue, whereas Lightning
fails dramatically (Nelson et al. 1979). Following Kintsch
(1988), we assumed that both the context cue Screw and its
target Bolt activate their associates during study, and that
this broad activation collapses to a specific meaning provid-
ed by the semantic context implied by the word pairing.
Because the target’s meaning is specified during study in
intralist cuing, its associates are no longer needed to support
its semantic state, and as a consequence, the density of the
target’s semantic neighborhood has no effect. The cue–tar-
get pairing can be directly loaded into working memory,
where it can be rehearsed and encoded. Furthermore, in
intralist cuing, collapse of activation is observed regardless
of whether or not a target word has multiple, distinct mean-
ings, because its meaning is more certain than when the
word is semantically isolated (Nelson et al., 1979).
Three findings reinforce the collapse assumption. First, in
intralist cuing, the target is studied in the presence of a se-
mantically related context word thatis used as the test cue, and
effects related to its associates are mostly eliminated (Holley
&McEvoy,1996; Nelson, Gee, & Schreiber, (1992), Nelson,
Schreiber, & McEvoy, (1992)). This collapse occurs when the
pairing is studied and is not produced by the test cue itself,
becausecollapseisobservedevenwhenrecallispromptedby
a rhyme cue linked to the target’s name (Nelson et al. 1992).
Second, the presence and timing of the semantic context word
in relation to its target is critical (Nelson et al. 1992). When the
context word appears slightly before or after the target, effects
related to the target’s associates are apparent once again, but
the extent of these effects falls between the levels during the
context-absent and the context-simultaneous conditions. Short
delays in the appearance of the context word introduce se-
mantic uncertainty and reveal the effects of the target’sasso-
ciates. Total collapse of the target’s associates depends on
simultaneously experiencing the cue–target pair. Finally, as
compared to when unrelated facts are memorized, facts orga-
nized around a semantic theme eliminate fan effects
(Anderson, 1983; Reder & Anderson, 1980)—that is, in-
creases in latencies with the number of a word’sassociates.
Fan effects are distractor effects, and as with word pairs,
encoding facts in a specific semantic context sharply reduces
distractor effects.
On the basis of these findings, we suspected that of the
four predictors in Eq. 1, only cue-to-target strength S
qt
should affect intralist cued recall, because the test cue must
be used to recover the cue–target pair in the semantic net-
work. In order to evaluate the effects of the four components
of Eq. 1on a larger scale, we rescored the data from 12
intralist-cuing experiments and created a database of 802
pairs. In the first analysis, the four variables were used as
predictors and the data were restricted to pairs presented at a
3-s rate and studied under instructions to remember the pair
and that the word appearing on the left would be the test cue.
The regression was significant, as is shown in Table 3, and
cue-to-target strength S
qt
is the only significant predictor. In
a second analysis, we included cue-to-target strength S
qt
,
presentation rate, and target frequency as predictors. The
regression was significant [F(3, 798) = 40.78, MS
res
= .035,
R= .37, R
2
= 13.0 %], as were all three predictors. For S
qt
,
presentation rate, and frequency, the coefficients were .31
(t= 9.24), .18 (t= 5.30), and –.14 (t=–4.15), and they
explained 8.4 %, 2.8 %, and 1.8 % of the total variance. The
recall means at study times of 1.5, 3.0, 5.0, and 6.0 s were
.75 (SE =.02,n= 93), .74 (SE = .01, n= 590), .81 (SE = .02,
n= 47), and .87 (SE = .02, n= 93).
As compared to extralist cuing, in intralist cuing, cue-to-
target strength and study time had much reduced effects, and
target-to-cue and cue-to-distractor strengths and, important-
ly, the target’s neighborhood density had no observable
effects. These contrasts indicate that the semantic network
responds dynamically to the context provided by the task at
hand. In extralist cuing, a word’s meaning as encoded dur-
ing study tends to be distributed across its semantic con-
texts, whereas in the intralist task, it tends to be confined and
specific, which comes as no surprise (e.g., Tulving &
Thomson, 1973). What is most interesting is the extralist-
cuing result showing that the semantic network supplies
meaning when a word is experienced out of context and
delivers what it has activated to a working memory system
that encodes this information for future use. The semantic
network responds dynamically to context and to memory
systems that have evolved for encoding recent events. Next,
we will consider effects of neighborhood density in single-
item recognition.
Mem Cogn (2013) 41:797–819 809
Recognition
In single-item recognition, subjects study a long list of words,
followed by a second list that is usually twice as long and
includes equal numbers of studied and new words. During
study, subjects process the words in some way, and the second
list tests recognition by having the subjects separate the test
words into two categories, “old”and “new.”In this task, the
test cue is the target itself, so of the four variables used in cued-
recall studies, only neighborhood density is relevant. Nelson et
al. (2001) asked subjects to study words that varied in density
and printed word frequency (see also Fisher & Nelson, 2006;
Nelson et al., 1998). Theoretically, targets with greater density
would be more likely to be correctly recognized when “old”
and correctly rejected when “new.”This finding holds across
variations in frequency and differences in study instructions
(“rate concreteness,”“name vowels,”or “remember”).
To deepen our understanding of these results and to
control for the effects of correlations between the predictor
variables, we set up a recognition database that included the
targets from all of our experiments. Targets were represented
in one column, with four measures of recognition performance
(d',A', hits, and false alarms) and the predictor variables
represented in the remaining columns. Table 3shows the
results of a simultaneous multiple regression of the pooled
data and indicates that all of the manipulated variables signif-
icantly predicted d' (and the other measures). Printed frequen-
cy effects are not shown in the table, but as expected they had
a very strong effect, with a coefficient of –.36, SE =.031,
t=–17.13. The study instructions (rate concreteness, name
vowels, or remember), frequency, and neighborhood density
were significantly correlated with d' at r=.38,–.37, and .12,
and these correlations explained 50.6 %, 46.9 %, and 1.7 % of
the variance that could be explained. Although the effect was
small, higher levels of neighborhood density were positively
correlated with single-item recognition.
In the recognition task, word frequency explains more
variance than does neighborhood density. A similar analysis
on extralist cuing disclosed a reversed pattern. Of the
explained variance, frequency and target density explained
2.7 % and 49.1 %, respectively. In both tasks, study in-
structions explained about the same amount of variance, but
recognition was more affected by distinctive letter patterns
(e.g., Malmberg, Steyvers, Stephens, & Shiffrin, 2002;
Shiffrin & Steyvers, 1997), whereas extralist cuing was
more affected by neighborhood density and target-priming
effects. Both tasks rely on both nonsemantic and semantic
information, but the same information plays different roles
in the two tasks. In recognition, the target is used as a cue for
itself, and its orthographic features play a more important
role than its semantic neighborhood. Alternatively, extralist
cuing relies on an external cue, and thus is forced to rely on
the semantic network. Recognition appears to depend more
on matching the target’s name, whereas extralist cuing de-
pends more on searching the network (e.g., Humphreys,
Bain, & Pike, 1989).
Summary of neighborhood density effects
The experimental and correlational findings indicate that
neighborhood density effects linked to the target facilitate
target recall and recognition to varying degrees. As density
increases, words are more likely to be recognized as studied
and to be rejected when not studied, and they are more likely
to be recalled when prompted by extralist test cues. In all of
these tasks, a target is experienced in semantic isolation
during study, and semantic memory responds by activating
its semantic neighborhood. Observing density effects ap-
pears to be contingent on both the absence of semantic
context during study and the presence of a related extralist
cue or a self-cue that provides access to its primed and
encoded representation in semantic memory.
The findings also show that the magnitudes of neighbor-
hood density effects in extralist cuing do not interact with
study time or level-of-processing manipulations. Density
effects are manifested at short and long study times, and
regardless of whether subjects are naming vowels, rehearsing,
or rating semantic attributes. Though the effects are small,
density also influences target recovery in primed free associ-
ation, where the importance of the study context and the cue–
target relationship are never mentioned. Finally, although
target recall declines quickly as a function of test delays, target
density effects are reduced but still apparent after 24 h.
However, target density effects are not evident when targets
are presented simultaneously with a semantically related word
that specifies their meanings, and though this can be
interpreted as indicating that the neighbors are not activated,
the speeded decision literature suggests otherwise (e.g.,
Kintsch, 1988): The target’sassociatesappeartobeactivated,
but this activation rapidly collapses because the associates are
irrelevant when target meaning is specified by the context.
The processes underlying neighborhood density effects, as
well as their collapse, appear to be automatic, unconscious,
and fast. Activation affords rapid access to prior knowledge in
the semantic network and is most relevant when the prevailing
semantic context is nonexistent or ambiguous. In order to
comprehendwhat it is experiencing, the brain rapidly provides
potentially relevant semantic contexts. Although activation is
normally treated in the speeded decision literature as an
ephemeral event lasting milliseconds, we suggest that neigh-
borhood density effects survive in semantically uncertain
contexts because primed targets are encoded in working
memory. A target’s primed representation functions as a foun-
dation for encoding processes performed upon it, and such
encoding ensures longer-lasting priming effects. The question
of why the semantic system operates in this way seems
810 Mem Cogn (2013) 41:797–819
straightforward: Given semantic uncertainty, activating prior
knowledge allows the system to quickly grasp the meaning of
current events in the broadest possible sense, in order to
anticipate unknown specific responses to future events.
When a future event arrives in the form of a retrieval cue, a
primed and encoded target is more likely to be selected within
the test cue’s noisy set of related distractors.
Sources of neighborhood density effects and target
priming
The most important question about the effects of neighbor-
hood density has been postponed, and it concerns the un-
derlying process: When the target is studied in semantic
isolation, how does the relative density of the links com-
prising its neighborhood affect its recall and recognition?
We assume that neighborhood density primes the represen-
tation of the target to greater or lesser degrees, but how does
the activation of the target’s associates prime the target?In
this section, we offer two explanations: spreading activation
and quantum-like entanglement.
Spreading activation
Initially, we turned to the Collins and Loftus (1975)spreading-
activation explanation. When applied to target-priming ef-
fects, spreading activation implies that seeing the target
spreads activation to and among its associates, and then back
to the target through associate-to-target links (Nelson et al.,
1998). Spreading activation is inherently a sequential process,
whereby activation moves from one word to the next along the
links in the network. To illustrate this idea, Fig. 6represents
target Twith links to two associates, one associate-to-associate
link, and one associate-to-target link. In the two-step loop,
activation spreads from Tto Associate 2 and back to T.Inthe
three-step loop, it travels to Associate 1, to Associate 2, and
then to T. According to this hypothesis, the effect of the link
from Associate 1 to Associate 2—or, more generally, the
effects of associate-to-associate links—arises because they
direct activation back to the target, rather than to words
outside its neighborhood.
According to spreading activation, target-priming effects are
contingent on the probability of links back to the target, and this
contingency predicts that effects of associate-to-associate links
must be greater when more associate-to-target pathways exist.
While this account is reasonable, the results of several experi-
ments that manipulated the proportions of both types of links in
a factorial design rejected this alternative (Nelson, McEvoy, &
Pointer, 2003). They showed that associate-to-associate effects
are robust as compared to associate-to-target effects, and more
importantly, that the expected interaction between the two
variables does not emerge; the target-priming effects of
associate-to-associate links are just as apparent when there
are few associate-to-target links as when there are many. In
addition, the preliminary regression analysis of extralist cuing
showed that associate-to-target links are not significantly
correlated with recall when other variables are controlled
(Table 2). Moreover, Bruza, Kitto, Nelson, and McEvoy
(2009) pooled the data from the experiments above and found
that a spreading-activation equation was a poor predictor of
recall (r= .35) as compared to Pier2 (r= .57). Hence,
experiencing a word does not appear to activate its nearest
neighbors sequentially. We assume that neighbors are activat-
ed in parallel, but this assumption begs the question of how
such an event can be modeled, and we turned to the concept of
quantum entanglement for an alternative explanation.
Semantic entanglement
The mathematics of quantum mechanics is being used to
generate a variety of models in the social sciences, and the
purpose of this section is to show how it can be used to model
target priming in extralist cuing (Aerts, 2009;Busemeyer&
Bruza, 2012; Busemeyer, Pothos, Franco, & Trueblood, 2011;
Gabora & Aerts, 2002; Pothos & Busemeyer, 2009).
Entanglement is a fundamental concept in quantum mechan-
ics that arises when two physical entities (such as electrons
and photons) are subjected to “preparations,”or what
cognitivists might call “procedures,”and these preparations
produce highly correlated experimental findings. Once the
entities are entangled, measuring the state of one of them
determines its state and simultaneously provides information
about the state of the other entity, even when it is so distant
that communication between them is impossible (see Bruza et
al., 2009, for more detail on how quantum theory is applied to
primed target strength; Greenstein & Zajonc, 2006,foran
introduction to experimental work; Laloe, 2001,forareview
of quantum theory; Maudlin, 1994, for a discussion of entan-
glement and its implications;and Turvey & Moreno, 2006,for
a rationale behind using physical metaphors to understand the
semantic network).
Fig. 6 Target Tactivates target-to-associate, associate-to-associate,
and associate-to-target links
Mem Cogn (2013) 41:797–819 811
Once entangled, the entities can be represented as a
“superposition state,”illustrated in the following equation
that describes a situation that arises before the spin states of
two particles are measured. Particle 1 has a potential of
spinning “down”while Particle 2 is spinning “up,”or,
alternatively, Particle 1 has a potential of spinning “up”
while Particle 2 is spinning “down”:
ψ
before measurement¼1
ffiffi2
p#1"2
ðÞþ"
1#2
ðÞ½
:
This superposition represents an indefinite state, and the
actual spin states of two entangled electrons cannot be
known until spin direction is measured. Measuring the state
of one of the electrons causes it to collapse to spinning “up”
or “down.”For example, if Particle 1 is measured, the state
could collapse to one state or the other:
ψ
after measurement ¼1
ffiffiffi
2
p#1"2
ðÞ½
or
ψ
after measurement ¼1
ffiffiffi
2
p"1#2
ðÞ½:
The important point is that measuring the state of Particle
1 predicts the state of Particle 2. In the “after”measurement
state, if Particle 1 is observed as spinning down, we know
that Particle 2 is spinning up, regardless of its distance from
Particle 1. Alternatively, if Particle 1 is observed as spinning
up, we know that Particle 2 will be observed as spinning
down. Once entangled, measuring the state of one particle
determines its state and, simultaneously, predicts the state of
the other particle.
To say the least, words differ from physical entities, but
certain abstract similarities make the comparisons worth
considering. For example, in both scientific disciplines en-
tities become entangled in some context, and measuring the
state of one of them simultaneously provides information
about its state and the states of correlated entities. In our
case, observing target-priming effects reveals the activation
states of a target’s associates. Our findings show that links
among the target’s associates prime its representation in
accordance with its neighborhood density. They also sug-
gest that neighborhood density effects are contingent on
parallel activation. An entangled state has to be “prepared,”
and this state emerges when the target’s neighborhood in the
semantic network is simultaneously activated. Theoretically,
simultaneous activation generates an indefinite superposi-
tion state in the experimental context, and quantum-like
entanglement provides a plausible hypothesis for explaining
how neighborhood density effects emerge from links among
its associates (Bruza et al., 2009; Busemeyer & Bruza, 2012;
Galea, Bruza, Kitto, Nelson, & McEvoy, 2011).
The preparation phase occurs during study, when the target
appears on the screen. Its appearance puts the target’snetwork
representation into a superposition state with its associates.
The model assumes that each of the associates is activated
with a probability of 0.0 and that each of the associates is
activated with a probability of 1.0. A word’s representation is
recovered from the network in an indefinite state in which each
of its associates is not and is activated. This state collapses to
one activation state or the other when it is observed in context,
and theoretically, this occurs when the representation appears
in working memory, where it is observed, read aloud, and
rehearsed, or rated on some attribute. The superposition col-
lapses to a definite state of being either not primed or primed
when the observer becomes aware of the target’sexistencein
the experimental context. Becoming aware of a word in the
setting is analogous to “measuring”the state of an electron, in
the sense that this measurement produces a definite state.
Entanglement model details
We will first illustrate how target priming is computed in
the entanglement model, and then describe the underlying
conceptualization. Table 7presents a matrix representa-
tion of target Tand its three associates. It shows how the
probability of activating each associate is calculated in
the bottom row, and how these associates affect target
priming. Theoretically, each probability measures the de-
gree to which an associate is activated by links from T’s
other associates, and Eq. 14 below shows how the en-
tanglement model combines the probabilities and predicts
target priming. Note that the equation indicates that each
associate contributes to priming, regardless of its seman-
tic distance from the target (readers may wish to skip
ahead to superposition Eq. 11 after examining this table).
The model treats the associates as “qubits”in a many-
bodied system. A qubit is a “quantum bit”and is a quantum
analogue of a classical bit. In classical probability theory, a
bit takes on a definite value of 0 or 1, whereas a qubit exists
as a weighted sum of both of these states that defines the
superposition. A classical bit represents information in a
binary fashion on the basis of logic states of “off”or “on.”
In contrast, a qubit can represent multiple states simulta-
neously for brief periods, because both alternative outcomes
are concurrently represented, along with a weighting term
that relates to their probability of occurring.
In quantum theory, “off”and “on”are represented in kets |0〉
and |1〉, which are called basis states that represent the context
of a quantum measurement.
2
Hence, the superposition state of a
system ψis written
2
The symbol |〉is called a “ket.”In quantum mathematics, kets provide
a context for describing probabilities—for instance, |0〉and |1〉are read
as “ket 0”and “ket 1.”
812 Mem Cogn (2013) 41:797–819
ψ
ji
¼a0
ji
þb1
ji
;
where αand βrepresent “probability amplitudes.”When α
and βare squared, they become probabilities; for instance, α
2
is the probability that a value of 0 is found, and β
2
is the
probability that a value of 1 is found. In quantum probability
theory, the indefinite superposition state collapses to one of
the specified states when a “measurement”is applied. In our
case, the activation state of a target is indefinite until it comes
into awareness, which causes its superposition state to col-
lapse to the “not activated”|0〉or the “activated”|1〉state.
Figure 7illustrates the system of related associates shown
in Table 7as a set of three qubits. Each qubit provides a spatial
representation for one of the target’s associates, represented as
a vector of length 1.0 for each associate. The probability of a
qubit collapsing to one state or the other is related to the
squared projection of the qubit state (the vector) down to the
basis state defined by the horizontal axis. This projection can
be envisioned as drawing a perpendicular line from the tip of
the vector onto the basis state. In quantum theory, probabilities
are based on the Pythagorean theorem a
2
+b
2
=c
2
,sotaking
the square root of the probability amplitude produces the
probability. Thus,the probability of an activated target is equal
to the square root of the projection onto the |1〉basis state, and
the probability of a nonactivated state is equal to the square
root of the projection to the |0〉basis state. These ket states take
the context of activation into account, so that the model is
inherently contextual, because the probability of activation
changes with changes in the context (Isham, 1995). Making
use of this procedure, we can write a superposition state for
each of the target’sassociatesa
1
,a
2
,anda
3
as
a1
ji
¼ffiffiffiffiffiffi
pa1
p0
ji
þffiffiffiffiffiffi
pa1
p1
ji
;ð8Þ
a2
ji
¼ffiffiffiffiffiffi
pa2
p0
ji
þffiffiffiffiffiffi
pa2
p1
ji
;ð9Þ
a3
ji
¼ffiffiffiffiffiffi
pa3
p0
ji
þffiffiffiffiffiffi
pa3
p1
ji
;ð10Þ
where a bar over pindicates “not activated.”
We take all three states in Eqs. 8–10 and combine them by
using a tensor product to produce the indefinite superposition
state. Such operations result in a complete set of possible
states (i.e., |000〉,|100〉….|111〉). However, because of our
assumption that either all or none of the associates are activat-
ed, the only states that can occur are |1〉⊗|1〉⊗|1〉=|111〉and
|0〉⊗|0〉⊗|0〉=|000〉, respectively. Given this restriction, the
activation state of the whole system can be written as an
entangled state:
ψ
t
ji
¼ffiffiffiffiffi
p0
p000
ji
þffiffiffiffiffi
p1
p111
ji
:ð11Þ
Equation 11 describes an entangled state representing the
target in its indefinite form, where p
0
and p
1
represent the
probabilities that none (|000〉)orthatall(|111〉)ofthe
associates are activated, respectively. We can extract values
for p
0
and p
1
by making the following observations using
the law of total probability:
p0¼pa1 pa2 pa3;ð12Þ
Table 7 Matrix representation of Target Twith three associates a, the links among them (reading along each row), and the probabilities (p) that an
associate will be activated
Target a
1
a
2
a
3
a
1
–11
a
2
1–1
a
3
00–
Probability p
a1
= (1 + 0) ÷ 3(3 –1) = .167 p
a2
= (1 + 0) ÷ 3(3 –1) = .167 p
a3
= (1 + 1) ÷ 3(3 –1) = .3333
Fig. 7 System of related
associates shown in Table 7,
represented as a set of three
qubits
Mem Cogn (2013) 41:797–819 813
and therefore,
p1¼1p0¼1pa1pa2pa3:ð13Þ
Given this activation assumption, the remaining proba-
bility mass contributes to primed target strength as a whole.
Consequently,
p1¼11pa1
ðÞ1pa2
ðÞ1pa3
ðÞ;ð14Þ
so in Table 7,
p1¼11:167ðÞ1:167ðÞ1:333ðÞ¼:537:
Equation 14 can be generalized for targets having any
number of associates, and we identify it and its assumptions
as the “entanglement model,”or EMS
t'
, for predicting the
target-priming effects described by S
t'
(Eq. 4). EMS
t'
is
successful to the extent that it correlates with S
t'
, and to
evaluate the model, we used Eq. 14 to compute the primed
target strength for all targets studied in the standard condi-
tions in the extralist-cuing database. The correlation between
EMS
t'
and S
t'
is r=.987,n=2,803,andR
2
=97.4%,sonearly
all of the variance of our intuitive descriptive statistic is
predicted by the entanglement model. When we enter EMS
t'
in place of S
t'
along with S
qt
,S
tq
,andS
qd
as predictors of recall
in a multiple regression analysis, the results are nearly identi-
cal [F(4, 2798) = 390.97, MS
res
=.043,R=.599,R
2
=.358],as
compared to when S
t'
is used (Table 3). The correlations with
probability of recall for EMS
t'
and S
t'
are r= .264 and .256,
respectively. Because the results are so similar, Eq. 1predicts
that the two methods for computing target activation will have
similar correlations with recall and similar predictions of mean
recall, and this is the case. Recall probability is correlated
r= .567 for both the EMS
t'
and S
t'
measures, and mean recall
is .392 (SD = .159) and .390 (SD = .162). The mean differences
between S
t'
and EMS
t'
tend to be very small, averaging .01,
because EMS
t'
tends to predict lower values of target activation
than does S
t'
for a few targets that have only several associates
(as in the example calculation in Table 7).
The entanglement model is successful because it captures
the effects of links among the target’s associates, as well as the
intuitive measure S
t'
. According to the model, the appearance
of the target during study puts it into an indefinite superposi-
tion state in which its associates are simultaneously activated
and not activated in the experimental context until the target’s
representation enters working memory, where it is observed
and can be rehearsed or its properties can be judged. As soon
as the target appears in working memory, subjects become
consciously aware of its existence in the experimental setting,
and the superposition state collapses to eithera nonprimed or a
primed state. This interpretation assumes that a minuscule
amount of time elapses between when a word appears on the
screen and when it is observed. During this interval, the target
is retrieved from the semantic network as a superposition
state. When the superposition collapses to its primed state,
priming effects emerge, and encoding a target in this state in
working memory links it to the contextual cues available in
the setting. In this approach, encoding processes applied in
accordance with the study instructions operate on the target’s
representation in working memory, regardless of the value of
its primed state in the system. When all of the target’sassoci-
ates are activated, a primed representation of Tis loaded into
working memory, and when none are activated, an unprimed
representation of Tis loaded into this system.
Another similarity that makes the comparison interesting
between physical and semantic states of entanglement con-
cerns the effects of semantic context. Quantum research has
indicated that superposition states are not permanent. Particles
and other objects existing in entangled states suddenly col-
lapse to one or the other state when measured (Laloe, 2001).
Analogously, seeing a target in semantic isolation locates it in
its semantic neighborhood and creates a superposition state in
the experimental setting that collapses to activating none or all
of its associates when a subject becomes consciously aware of
it in working memory. In contrast, when the target is seen in
the context provided by a semantically related word, neigh-
borhood density effects are not apparent, because the target’s
meaning is specified by the context. Seeing a semantically
related cue–target pairing creates an entangled state in which
the target’s associates are or are not activated, but this state
immediately collapses to the nonactivated state when this
representation reaches working memory. The target’ssuper-
position state prior to appearing in working memory is the
same as when it is seen in isolation, but when it is seen in an
appropriate semantic context, the superposition state collapses
to a nonactivated state when observed in working memory.
Although derived from an unusual source for the field of
memory, the entanglement model fares better as a model of
target activation effects than does spreading activation, be-
cause it is based on a parallel activation process that is sensi-
tive to environmental context.
General discussion
Both experimental and correlational findings have shown
how the semantic network contributes to event memory in
tasks involving a semantically isolated target and a cue that
is related to the target, or that is the target in a physical
sense. The magnitude of the network contribution based on
variance explained differs across tasks, depending on the
semantic context during encoding, context awareness during
testing, and the nature of the cue. Importantly, in extralist
cuing the network contributes more to recall than does study
time or how the target is encoded during study. Recall in this
814 Mem Cogn (2013) 41:797–819
task is affected by both semantic and episodic information,
and the semantic information is critical, because recall is
based on a test cue that is both absent during study and
semantically related to the target. The network provides
semantic information during study that defines the meaning
of an isolated word event, and during testing it provides
information that guides the search for this event. During
both study and test, familiar words act as reminding cues
that reliably and quickly activate the prior semantic contexts
of these words. In what follows, we consider four general
and partially overlapping conclusions.
The semantic network
The first conclusion is that a semantic network exists in
memory as a complex system of words and links that func-
tions as a repository of word knowledge. Words are not
randomly connected in this system (Collins & Loftus,
1975; Steyvers & Tenenbaum, 2005). Rather, they cluster
around similar experiences involving thousands of direct
links and tens of thousands of indirect links that may pro-
vide shortcuts across the domain (Watts & Strogatz, 1998).
This system continually interacts with familiar words, ob-
jects, and events by automatically activating related words.
Automatic parallel activation provides the informational
bases for comprehension, priming, and encoding, as well
as for search, recall, and recognition. Episodic encoding and
retrieval depend to varying degrees on the semantic network
to provide the word knowledge needed in each of these
processes, and more specifically, the findings show that
measures based on the network are correlated with memory
performance in several event memory tasks. Of the four key
measures taken on the network, word recognition relies only
on target activation; intralist cuing relies only on cue-to-
target strength; and extralist cuing depends on all four
measures. Taken together, the findings provide strong evi-
dence for the hypothesis that words are represented and
processed in a vast dynamic semantic storage system that
contributes to and that can be used to predict the recall and
recognition of episodic events.
Free association and semantic context
The second general conclusion is that free association to
thousands of words can be used to generate a semantic
network that measures what related words are likely to come
to mind when experiencing a word—measures that are
reliable and valid, and that are sensitive to semantic context.
This method is reliable because renorming the same word
tends to produce the same associates for even weaker asso-
ciates, and it is valid because measures taken from this
network predict behavior in various memory tasks. Finally,
the network is sensitive to context, because the same word
takes on different shades of meaning when influenced by
different cues. For example, in our free-association norms,
food is produced as a response by 324/5,018 different cues,
and its meaning varies for each cue. Each cue that produces
food as an associate also produces related words that vary
substantially across cues, so that food means something
different in the context of different words. Both Dinner
and Hamburger produce food as an associate, but otherwise
their associates are completely different, indicating that food
activates different semantic senses in the contexts provided
by Dinner and Hamburger. Free-association norms capture
context-driven differences in meaning, because a word with
exactly the same name appears among different associates
for each cue that generates it in free association. We suggest
that the semantic network generated via free association is
highly sensitive to semantic context differences because the
cues being normed activate not just a single word, but an
array of related words out of which one or more words can
be selected.
Activation, search, and memory task
The third conclusion concerns the distinction between the
activation associated with priming effects and search. The
most advanced memory models of recall generally do not
explicitly incorporate activation in the modeling, because
the research is based on the free recall of unrelated words,
defined as words that are not directly associated (e.g.,
Gillund & Shiffrin, 1984; Shiffrin & Steyvers, 1997). The
study phases in free-recall and extralist-cuing tasks are
essentially identical, so it is reasonable to assume that the
target words are loaded into working memory as either
primed or not primed, but the testing phases are different.
In free recall, subjects are asked to recall all list words in any
order, whereas in cued recall, a semantically related cue is
used to prompt the recall of each one. In extralist cuing, the
semantic cue provides access to an array of response alter-
natives that include the target, whereas in free recall the first
word recalled is cued by context, and the remaining words
are cued through a combination of context and episodic
associations between the words produced by rehearsal dur-
ing study (Raaijmakers, & Shiffrin, 1981). In the free recall
of unrelated words, what is activated in the semantic net-
work serves no apparent purpose. There is no reason to
believe that the activation of related words by the target or
the cue might be important, so the priming literature and the
concept of activation have been largely and reasonably
ignored in modeling free recall.
The results of three experiments that compared the two
tasks directly, using the same materials and study conditions,
justified this position (Nelson, Akirmak, & Malmberg, in
preparation). These experiments showed that neighborhood
density and the resulting priming effects are strong in extralist
Mem Cogn (2013) 41:797–819 815
cuing and absent in free recall. Although both tasks involve
search, the sets that are being searched are different: In free
recall, memory for the study list is searched using a recalled
word as a cue for recalling another word from the study list, on
the basis of an episodically created link (Raaijmakers &
Shiffrin, 1981). In contrast, in extralist cuing, the cue activates
its associates in the semantic network, and search is limited to
this set. The network becomes involved in recall when a test
cue is semantically linked to its prospective target, and the task
comparison suggests that target-priming effects are contingent
on recovering the target from its storage in this network:
Target-priming effects may be observed only when searching
the network that created these effects in the first place.
These findings suggest that the network plays a critical role
in recall that is cued by a meaningfully related word, but no
role whatsoever when the recall of one list word is cued by an
unrelated word. They also indicate that, when the semantic
system is utilized, effects of related words can be apparent
during both study and test. Related words are not physically
presented, so some brain mechanism must be providing them,
and the findings here and in the priming literature suggest that
implicit activation is that mechanism. The evidence reviewed
here suggests that the semantic memory system supports the
recognition and recall of episodic events by activating related
words that serve as the basis for comprehension and search
processes in some, but not all, episodic memory tasks. In the
extralist-cuing task, the simultaneous activation of the target’s
associates during the study phase provides the semantic con-
text for a word seen in isolation, and during testing, the cue
activates a set of potential responses that is likely to include
the target and that initiates a search for the target within this
set. Activation plays a role in both phases, but the study phase
is largely concerned with comprehension and following the
study instructions. The key network variable in this phase is
neighborhood density—specifically, its influence on target
priming. In contrast, the test phase is largely concerned with
search, and the key network variables are cue-to-target
strength, cue-to-distractor strength, and target-to-cue strength.
Variables associated with the semantic network contribute in
different ways to both phases of extralist cuing, and in differ-
ent ways to recognition and intralist cuing, suggesting that the
network responds dynamically to the contexts provided by
different ways of measuring memory. Semantic memory is a
fast-acting dynamic system that reminds us about what we
know. To return to the beginning, our findings indicate that the
electric company is paid because check,envelope,andmail-
box activate related words that allow mailbox to succeed as a
cue by reminding us to pay the bill.
Pier3 and entanglement
Equation 1provides an algorithm for predicting extralist cuing
that evolved from the Pier framework, and we refer to it as
Pier3. Pier3 computations are derived from measures made on
the semantic network that are combined in a ratio rule that
predicts the contribution of the semantic system to probability
of recall. Pier3’s core rationale is motivated by the SAM
model (“search of associative memory”;Raaijmakers&
Shiffrin, 1981), but it is incomplete as a SAM model, because
it does not incorporate the effects of environmental context or
of encoding operations in calculating its predictions. In addi-
tion, it differs from SAM in that search is based on a semantic
network of links acquired in interactions with the world, as
opposed to an episodic network of associative links
established during recent study. In addition, Pier3 makes
probabilistic predictions for specific cue–target pairs. By com-
puting mean predicted recall and the standard deviation for
pairs linked to a specific semantic manipulation (e.g., neigh-
borhood density), it predicts the contribution of the semantic
network to recall in extralist-cuing experiments. In extralist
cuing, once a list is prepared, it is possible to compute a recall
prediction for each cue–target pair, and by averaging over the
pairs in a given condition, it is possible to predict recall for that
condition. Finally, Pier3 also differs from SAM in how it
aggregates links among related words, but despite the differ-
ences, both approaches assume that recall is based on
searching memory for recently encoded information. Given
this commonality, it might be possible to develop a SAM
model for extralist cuing that incorporates simulations of
episodic encoding processes using predictions based on se-
mantic relationships. In Bayesian terms, Pier3 predictions
could serve as the prior probability, and a to-be-developed
measure of the contribution of episodic memory could serve
as the posterior probability.
In Pier3, the test cue initiates a search of the semantic
network, and this search is facilitated when the cue and its
target are more strongly linked relative to all of the informa-
tion relevant to its selection, including the error introduced by
the cue’s distractors. Recall is more likely when the cue-to-
target and target-to-cue strengths are high, and when cue-to-
distractor strength is weak. It is also more likely when target
priming is higher, when the target is studied longer, and when
it is studied with a more effective procedure. Pier3 assumes
that both target priming and encoding strengthen its represen-
tation in the semantic network and that this elevation decays
over time unless refreshed.
The target-priming effect is in many ways a most inter-
esting phenomenon, because it is founded on links among
the target’s associates. Classic spreading-activation theory
predicts that this result arises because activation in the
target’s network returns to the target, as opposed to spread-
ing outside its neighborhood. This interpretation, however,
is countered by evidence indicating that target-priming ef-
fects are independent of the presence of associate-to-target
links. Target-priming effects do not appear to arise from a
sequential activation process, but from a parallel process
816 Mem Cogn (2013) 41:797–819
that simultaneously activates the target’s neighborhood.
Words exist in a tangled network of associative/semantic
links, and when context fails to provide a specific semantic
sense, the network activates and temporarily encodes its
most immediate senses. The entanglement model captures
this process effectively because it is based on an indefinite
superposition state that is sensitive to context, represented in
the probabilities that none and all of a target’sassociatesare
activated. The model assumes that the physical appearance of
a word puts it into a superposition state that collapses to a
specific sense when the word reaches working memory, and
an individual becomes aware of its existence. The target
becomes entangled with its associates, and such entanglement
occurs when the target and its associates are simultaneously
activated. When applied to the extralist-cuing database, the
model explains 97.4 % of the variance of the intuitive measure
of neighborhood density. Although rarely used in psycholog-
ical research, this result indicates that the superposition as-
sumption can be productively used to study difficult problems
involving indefinite semantic states and, possibly, other indef-
inite psychological states involving uncertainty (Busemeyer
&Bruza,2012).
Author note The authors thank Lili Sahakyan and the reviewers and
editor for their exceptionally helpful suggestions. We also offer heart-
felt thanks to all of the students who worked on this project over its 30-
year existence. This project was supported in part by Australian Re-
search Council Discovery grant DP1094974.
References
Aerts, D. (2009). Quantum structure in cognition. Journal of
Mathematical Psychology, 53, 314–348.
Anderson, J. R. (1983). A spreading activation theory of memory.
Journal of Verbal Learning and Verbal Behavior, 22, 261–295.
doi:10.1016/S0022-5371(83)90201-3
Barabási, A.-L. (2002). Linked: The new science of networks.
Cambridge, MA: Perseus.
Battig, W. F., & Montague, W. E. (1969). Category norms of verbal items
in 56 categories: A replication and extension of the Connecticut
category norms. Journal of Experimental Psychology, 80(3, Pt. 2),
1–46. doi:10.1037/h0027577
Borge-Holthoefer, J., & Arenas, A. (2010). Semantic networks: struc-
ture and dynamics. Entropy, 12, 1264–1302.
Brainerd, C. J., Yang, Y., Reyna, V. F., Howe, M. L., & Mills, B. A.
(2008). Semantic processing in “associative”false memory.
Psychonomic Bulletin & Review, 15, 1035–1053. doi:10.3758/
PBR.15.6.1035
Bruza, P., Kitto, K., Nelson, D., & McEvoy, C. (2009). Is there
something quantum-like about the human mental lexicon?
Journal of Mathematical Psychology, 53, 362–377. doi:10.1016/
j.jmp.2009.04.004
Busemeyer, J. R., & Bruza, P. (2012). Quantum models of cognition and
decision making. New York, NY: Cambridge University Press.
Busemeyer, J. R., Pothos, E., Franco, R., & Trueblood, J. (2011). A
quantum theoretical explanation for probability judgment errors.
Psychological Review, 118, 193–218.
Chappell, M., & Humphreys, M. S. (1994). An auto-associative neural
network for sparse representations: Analysis and application to models
of recognition and cued recall. Psychological Review, 101, 103–128.
Collins, A. M., & Loftus, E. F. (1975). A spreading-activation theory
of semantic processing. Psychological Review, 82, 407–428.
doi:10.1037/0033-295X.82.6.407
Cramer, P. (1968). Word association. New York, NY: Academic Press.
De Deyne, S., & Storms, G. (2008a). Word associations: Network and
semantic properties. Behavior Research Methods, 40, 213–231.
doi:10.3758/BRM.40.1.213
De Deyne, S., & Storms, G. (2008b). Word associations: Norms for
1,424 Dutch words in a continuous task. Behavior Research
Methods, 40, 198–205.
Deese, J. (1965). The structure of associations in language and
thought. Baltimore, MD: Johns Hopkins Press.
Fisher, S. L., & Nelson, D. L. (2006). Recursive reminding: Effects of
repetition, printed frequency, connectivity, and set size on recog-
nition and judgments of frequency. Memory & Cognition, 34,
295–303. doi:10.3758/BF03193408
Gabora, L., & Aerts, D. (2002). Contextualizing concepts using a math-
ematical generalization of the quantum formalism. Journal of
Experimental and Theoretical Artificial Intelligence, 14, 327–358.
Galea, D., Bruza, P. D., Kitto, K., Nelson, D., & McEvoy, C. L. (2011).
Modeling the activation of human memory: The spreading acti-
vation, spooky action at a distance, and the entanglement models
compared. In D. Song & M. Melucci (Eds.), Proceedings of the
5th Quantum Interaction International Symposium (QI-2011) vol.
7052 (pp. 149–160). Berlin: Springer.
Gee, N. (1997). Implicit memory and word ambiguity. Journal of
Memory and Language, 36, 253–275.
Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both recog-
nition and recall. Psychological Review, 91, 1–67. doi:10.1037/
0033-295X.91.1.1
Goldman, M. S. (1999). Risk for substance abuse: Memory as a common
etiological pathway. Psychological Science, 10, 196–199.
Goodmon, L. B., & Nelson, D. L. (2004). Strengthening the activation
of unconsciously activated memories. Memory & Cognition, 32,
804–818. doi:10.3758/BF03195870
Greenstein, G., & Zajonc, A. G. (2006). The quantum challenge:
Modern research on the foundations of quantum mechanics
(2nd ed.). Sudbury, MA: Jones & Bartlette.
Griffiths, T. L., Steyvers, M., & Tenenbaum, J. B. (2007). Topics in
semantic representation. Psychological Review, 114, 211–244.
doi:10.1037/0033-295X.114.2.211
Hills, T. T., Maouene, M., Riordan, B., & Smith, L. (2010). The
associative structure of language: Contextual diversity in
early word learning. Journal of Memory and Language, 63,
259–273.
Holley, P. E., & McEvoy, C. L. (1996). Aging and inhibition of
unconsciously processed information: No apparent deficit.
Applied Cognitive Psychology, 10, 241–256.
Humphreys, M. S., Bain, J. D., & Pike, R. (1989). Different ways to
cue a coherent memory system: A theory for episodic, semantic,
and procedural tasks. Psychological Review, 96, 208–233.
doi:10.1037/0033-295X.96.2.208
Humphreys, M. S., Tehan, G., O’Shea, A., & Bolland, S. W. (2000).
Target similarity effects: Support for the parallel distributed pro-
cessing assumptions. Memory & Cognition, 28, 798–811.
Hutchison, K. A. (2003). Is semantic priming due to association strength
or feature overlap? A microanalytic review. Psychonomic Bulletin &
Review, 10, 785–813. doi:10.3758/BF03196544
Isham, C. J. (1995). Lectures on quantum theory. London, UK:
Imperial College Press.
Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N., & Barabási, A.-L.
(2000). The large-scale organization of metabolic networks.
Nature, 407, 651–654.
Mem Cogn (2013) 41:797–819 817
Kintsch, W. (1988). The role of knowledge in discourse comprehen-
sion: a construction–integration model. Psychological Review, 95,
163–182. doi:10.1037/0033-295X.95.2.163
Laloe, F. (2001). Do we really understand quantum mechanics?
Strange correlations, paradoxes and theorems. American Journal
of Physics, 69, 655–701.
MacRae, K., Seidenberg, M. S., & McNorgen, C. (2005). Semantic
feature production norms for a large set of living and nonliving
things. Behavior Research Methods, 37, 547–559.
Maki, W. S., & Buchanan, E. (2008). Latent structure in measures of
associative, semantic, and thematic knowledge. Psychonomic
Bulletin & Review, 15, 598–603. doi:10.3758/PBR.15.3.598
Malmberg, K. J., Steyvers, M., Stephens, J. D., & Shiffrin, R. M.
(2002). Feature frequency effects in recognition memory.
Memory & Cognition, 30, 607–613. doi:10.3758/BF03194962
Maudlin, T. (1994). Quantum non-locality and relativity: Metaphysical
intimations of modern physics. Oxford, UK: Oxford University
Press.
Monaco, J. D., Abbott, L. F., & Kahana, M. J. (2008). Lexico-semantic
structure and the word frequency effect in recognition memory.
Learning and Memory, 14, 204–213.
Moreno, J. L. (1934). Who shall survive? Beacon NY: Beacon House.
Myers, J. L., & Well, A. D. (1995). Research design and statistical
analysis. Hillsdale: Erlbaum.
Nelson, D. L., Bajo, M.-T., & Casanueva, D. (1985). Prior knowledge
and memory: The influence of natural category size as a function
of intention and distraction. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 11, 94–105. doi:10.1037/
0278-7393.11.1.94
Nelson, D. L., Bajo, M.-T., McEvoy, C. L., & Schreiber, T. A. (1989).
Prior knowledge: The effects of natural category size on memory
for implicitly encoded concepts. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 15, 957–967.
doi:10.1037/0278-7393.15.5.957
Nelson,D.L.,Bennett,D.J.,Gee,N.R.,Schreiber,T.A.,&
McKinney, V. M. (1993). Implicit memory: Effects of network
size and interconnectivity on cued recall. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 19, 747–764.
doi:10.1037/0278-7393.19.4.747
Nelson, D. L., & Castano, D. (1984). Mental representations for
pictures and words: same or different? The American Journal of
Psychology, 97, 1–15.
Nelson, D. L., Dyrdal, G. M., & Goodmon, L. B. (2005). What is pre-
existing strength? Predicting free association probabilities, simi-
larity ratings, and cued recall probabilities. Psychonomic Bulletin
& Review, 12, 711–719.
Nelson, D. L., Fisher, S. L., & Akirmak, U. (2007a). How implicitly
activated and explicitly acquired knowledge contribute to the
effectiveness of retrieval cues. Memory & Cognition, 35, 1892–
1904. doi:10.3758/BF03192923
Nelson, D. L., Gee, N. R., & Schreiber, T. A. (1992a). Sentence
encoding and implicitly activated memories. Memory &
Cognition, 20, 643–654.
Nelson, D. L., & Goodmon, L. B. (2002). Experiencing a word can
prime its accessibility and its associative connections to related
words. Memory & Cognition, 30, 380–398. doi:10.3758/
BF03194939
Nelson, D. L., & Goodmon, L. B. (2003). Disrupting attention: The
need for retrieval cues in working memory theories. Memory &
Cognition, 31, 65–76. doi:10.3758/BF03196083
Nelson, D. L., Goodmon, L. B., & Akirmak, U. (2007b). Implicitly
activated memories are associated to general context cues.
Memory & Cognition, 35, 1878–1891. doi:10.3758/BF03192922
Nelson, D. L., Goodmon, L. B., & Ceo, D. (2007c). How does delayed
testing reduce effects of implicit memory: Context infusion or
cuing with context? Memory & Cognition, 35, 1014–1023.
doi:10.3758/BF03193474
Nelson, D. L., & McEvoy, C. L. (1979). Encoding context and set size.
Journal of Experimental Psychology: Human Learning and
Memory, 5, 292–314.
Nelson, D. L., McEvoy, C. L., & Dennis, S. (2000). What is free
association and what does it measure? Memory & Cognition, 28,
887–899. doi:10.3758/BF03209337
Nelson, D. L., McEvoy, C. L., & Pointer, L. (2003). Spreading activa-
tion or spooky action at a distance? Journal of Experimental
Psychology: Learning, Memory, and Cognition, 29, 42–51.
doi:10.1037/0278-7393.29.1.42
Nelson,D.L.,McEvoy,C.L.,&Schreiber,T.A.(2004).The
University of South Florida free association, rhyme, and word
fragment norms. Behavior Research Methods, Instruments, &
Computers, 36, 402–407. doi:10.3758/BF03195588
Nelson, D. L., McKinney, V. M., Gee, N. R., & Janczura, G. A. (1998).
Interpreting the influence of implicitly activated memories on
recall and recognition. Psychological Review, 105, 299–324.
Nelson, D. L., Schreiber, T. A., & McEvoy, C. L. (1992b). Processing
implicit and explicit representations. Psychological Review, 99,
322–348.
Nelson, D. L., Walling, J. R., & McEvoy, C. L. (1979). Doubts about
depth. Journal of Experimental Psychology: Human Learning
and Memory, 5, 24–44.
Nelson, D. L., & Zhang, N. (2000). The ties that bind what is known to
the recall of what is new. Psychonomic Bulletin & Review, 7, 604–
617. doi:10.3758/BF03212998
Nelson, D. L., Zhang, N., & McKinney, V. M. (2001). The ties that
bind what is known to the recognition of what is new. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 27,
1147–1159. doi:10.1037/0278-7393.27.5.1147
Newman, M. E. J. (2003). The structure and function of complex
networks. SIAM Review, 45, 167–256.
Palla, G., Derényi, I., Farkas, I., & Vicsek, T. (2005). Uncovering the
overlapping community structure of complex networks in nature
and society. Nature, 435, 814–818.
Pothos, E., & Busemeyer, J. R. (2009). A quantum probability expla-
nation for violations of “rational”decision theory. Proceedings of
the Royal Society B, 276, 193–218.
Raaijmakers, J. G. W., & Shiffrin, R. M. (1981). Search of associative
memory. Psychological Review, 88, 93–134. doi:10.1037/0033-
295X.88.2.93
Reder, L. M., & Anderson, J. R. (1980). A partial resolution of the
paradox of interference: The role of integrating knowledge.
Cognitive Psychology, 12, 447–472.
Reich, R. R., & Goldman, M. (2005). Exploring the alcohol expectan-
cy network: The utility of free associates. Psychology of Addictive
Behaviors, 19, 317–325.
Sahakyan, L., & Goodmon, L. B. (2010). Theoretical implications of
extralist probes for directed forgetting. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 36, 920–937.
Shiffrin, R. M., & Steyvers, M. (1997). A model for recognition memory:
REM—retrieving effectively from memory. Psychonomic Bulletin
& Review, 4, 145–166. doi:10.3758/BF03209391
Stacy, A. W. (1997). Memory activation and expectancy as prospective
predictors of alcohol and marijuana use. Journal of Abnormal
Psychology, 106, 61–73.
Steyvers, M., Shiffrin, R. M., & Nelson, D. L. (2005). Word associa-
tion spaces for predicting semantic similarity effects in episodic
memory. In A. F. Healy (Ed.), Experimental cognitive psychology
and its applications (pp. 237–249). Washington, DC: American
Psychological Association.
Steyvers, M., & Tenenbaum, J. B. (2005). The large-scale structure
of semantic networks: Statistical analyses and a model of
818 Mem Cogn (2013) 41:797–819
semantic growth. Cognitive Science, 29, 41–78. doi:10.1207/
s15516709cog2901_3
Strogatz, S. H. (2003). SYNC: The emerging science of spontaneous
order. New York, NY: Hyperion Books.
Tulving, E., & Thomson, D. M. (1973). Encoding specificity and
retrieval processes in episodic memory. Psychological Review,
80, 352–373. doi:10.1037/h0020071
Turvey, M. T., & Moreno, M. A. (2006). Physical metaphors for the
mental lexicon. Mental Lexicon, 1, 7–33.
Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of “small-
world”networks. Nature, 393, 440–442.
Zeelenberg, R., Shiffrin, R. M., & Raaijmakers, J. G. W. (1999). Priming
in a free association task as a function of association directionality.
Memory & Cognition, 27, 956–961. doi:10.3758/BF03201226
Mem Cogn (2013) 41:797–819 819
CopyrightofMemory&CognitionisthepropertyofSpringerScience&BusinessMedia
B.V.anditscontentmaynotbecopiedoremailedtomultiplesitesorpostedtoalistserv
withoutthecopyrightholder'sexpresswrittenpermission.However,usersmayprint,
download,oremailarticlesforindividualuse.
