Conference PaperPDF Available

Simple Mechanisms, Rich Structure: Statistical Co-Occurrence Regularities in Language Shape the Development of Semantic Knowledge


Abstract and Figures

Many hallmarks of human intelligence including language, reasoning, and planning require us to draw upon knowledge about the world in which concepts, denoted by words, are organized by meaningful, semantic links between them (e.g., juicy-apple-pear). The goal of the present research was to investigate how these organized semantic networks may emerge in development from simple but powerful mechanisms sensitive to statistical co-occurrence regularities of word use in language. Specifically, we tested whether a mechanistic account of how co-occurrence regularities shape semantic development accurately predicts how semantic organization changes with development. Using a sensitive, gaze-based measure of the semantic links organizing knowledge in children and adults, we observed that developmental changes in semantic organization were consistent with a key role for statistical co-occurrence regularities.
Content may be subject to copyright.
Simple Mechanisms, Rich Structure: Statistical Co-Occurrence Regularities in
Language Shape the Development of Semantic Knowledge
Layla Unger (
Olivera Savic (
Vladimir M. Sloutsky (
Ohio State University, Department of Psychology, 1835 Neil Avenue
Columbus, OH 43210 USA
Many hallmarks of human intelligence including language,
reasoning, and planning require us to draw upon knowledge about
the world in which concepts, denoted by words, are organized by
meaningful, semantic links between them (e.g., juicy-apple-pear).
The goal of the present research was to investigate how these
organized semantic networks may emerge in development from
simple but powerful mechanisms sensitive to statistical co-
occurrence regularities of word use in language. Specifically, we
tested whether a mechanistic account of how co-occurrence
regularities shape semantic development accurately predicts how
semantic organization changes with development. Using a
sensitive, gaze-based measure of the semantic links organizing
knowledge in children and adults, we observed that
developmental changes in semantic organization were consistent
with a key role for statistical co-occurrence regularities.
Keywords: semantic organization; semantic development;
statistical learning; taxonomic; association
We rely on our knowledge about the world to achieve many
vital, every-day cognitive tasks. For example, our knowledge
of apples can allow us to use language to express and
comprehend ideas about eating apples, retrieve knowledge
from memory that apples are healthy to achieve a goal to eat
a healthy snack, plan for lunch by packing an apple, and
generate new ideas such as “pears are healthy” by
generalizing what we know about apples to other fruits.
These feats rely on knowledge that is not a jumble of facts,
but instead an organized semantic network of linked
concepts, such as apples, pears, eating, and healthy. How
does such vital semantic organization emerge and change in
the course of development?
The many prior accounts of knowledge organization
development have focused on how we form taxonomic links
between members of the same, stable category, such as links
between pigeon and duck that belong to the category of birds
(Gelman & Markman, 1986; Inhelder & Piaget, 1964;
Lucariello, Kyratzis, & Nelson, 1992; Sloutsky, 2010).
Research into taxonomic link development may provide
valuable insights into the development of semantic
organization. However, semantic organization is much richer
than just these links, encompassing a variety of taxonomic
and non-taxonomic links between a multitude of concepts.
The goal of the present research is to evaluate how this richer
semantic structure may be driven in part by simple but
powerful mechanisms that form semantic links based on
statistical co-occurrence regularities of word use in language.
Co-Occurrence in Semantic Development
The potentially fundamental roles for co-occurrence that we
will investigate are outlined in the recently proposed Co-
Occurrence Account (Sloutsky, Yim, Yao, & Dennis, 2017).
According to this account, sensitivity to co-occurrence
initially fosters associative links between concepts whose
labels reliably occur close together in language (adjacent or
separated by intervening words), such as juicy-apple.
Henceforth, these regularities are referred to as direct co-
occurrence. These associative links form a key non-
taxonomic facet of semantic networks that supports
knowledge-dependent cognition from early development
onward. For example, upon hearing a new word such as dax
accompanied by words associated with animal such as furry,
both young children and adults infer that dax means animal
(Sloutsky et al., 2017)
Critically, sensitivity to co-occurrence can also foster
taxonomic links, because words for members of taxonomic
categories (e.g., “apple” and “pear”) reliably share
overlapping patterns of direct co-occurrence with other
words (e.g., "juicy", Jones, Willits, & Dennis, 2015)
However, unlike direct co-occurrence, shared co-occurrence
cannot be immediately gleaned from language input. For
example, to form a shared co-occurrence-based link between
apple and pear as shown in Figure 1, the learner must
Figure 1: Direct and shared co-occurrence regularities
that can form associative and taxonomic links.
experience direct co-occurrences between both juicy and
apple, and juicy and pear, then form a link between
apple and pear based on their overlapping direct co-
occurrence with juicy
. Therefore, taxonomic links may
develop more gradually (Bauer & Larkina, 2017; Schlichting,
Guarino, Schapiro, Turk-Browne, & Preston, 2017).
Together, the learning processes proposed in the Co-
Occurrence account outline how simple co-occurrence
regularities may build both associative and taxonomic links
between any concepts denoted by words. Moreover, this
account makes specific predictions about how these semantic
links emerge in the course of development. In what follows,
we describe evidence from prior research supporting this
account, then present an experiment designed to evaluate its
predictions about the development of semantic organization.
Support for the Co-Occurrence Account. The Co-
occurrence account is motivated by extensive evidence that
linguistic input is rich in regularities from which semantic
links can be formed. First, much of the variance in the
strength of semantic links in adult semantic networks can be
predicted by regularities with which words directly co-occur
or share co-occurrence in language (Hofmann, Biemann,
Westbury et al., 2018; Spence & Owens, 1990). Moreover,
computational models that form word representations based
on co-occurrence statistics in language simulate semantic
networks that predict complex semantic phenomena, from
semantic priming effects to the typical vocabulary growth
rate of school children (Jones et al., 2015; Landauer &
Dumais, 1997; Sahlgren, 2008). Together, these findings
ground the Co-occurrence account’s proposal that co-
occurrence regularities are important drivers of semantic
organization development.
Importantly, a key proposal of the Co-Occurrence account
is that developing humans form semantic links from co-
occurrence regularities in language. Extensive evidence
supports the possibility that humans form links based on
direct co-occurrence starting early in development. Much of
this evidence comes from statistical learning research, which
has shown early-developing abilities to form direct co-
occurrence-based links between stimuli such as speech
sounds (Saffran, Aslin, & Newport, 1996) and images (Fiser
& Aslin, 2002). Recently, the formation of direct co-
occurrence-based links between words has also been
observed in toddlers (Wojcik & Saffran, 2015) and young
children (Matlen, Fisher, & Godwin, 2015).
Evidence that humans form shared co-occurrence-based
links comes from a smaller body of research. For example,
adults in Preston, Zeithamova and colleagues’ studies
(Zeithamova, Dominick, & Preston, 2012) who explicitly
memorize pairs of images also link images that were never
paired, but instead share each other’s pairing with the same
image (see also Hall, Mitchell, Graham, & Lavis, 2003;
The mechanism(s) that form links between inputs that share
patterns of co-occurrence remain unknown. Candidates have been
proposed and investigated in multiple fields, including conditioning
(Honey & Hall, 1989), hippocampal memory formation (Schapiro,
Turk-Browne, Botvinick, & Norman, 2017), and semantic
Schapiro, Rogers, Cordova, Turk-Browne, & Botvinick,
2013). Moreover, this ability may develop only gradually.
For example, Schlichting et al. (2017) observed that the
ability to link images based on their shared pairing with
another image improved substantially from age six to
adulthood. Similar evidence for gradual development comes
from studies conducted by Bauer and colleagues, in which
participants were given two stem facts that both link
information to a shared concept, such as “dolphins talk by
clicking and squeaking” and dolphins live in groups called
pods. The ability to integrate across stem facts to derive a
new fact such as pods talk by clicking and squeaking is
poor at age four, and substantially increases over childhood.
This prior evidence motivates and supports Co-occurrence
account’s proposals about how sensitivities to direct and
shared co-occurrence regularities may build semantic
organization. However, evaluating the Co-occurrence
account critically involves testing the specific predictions it
makes about the development of semantic organization.
Specifically, the Co-Occurrence account predicts that
associative links between concepts that can form from direct
co-occurrence regularities should emerge early in
development. With development, associative links should
become gradually supplemented by taxonomic links that can
be formed from shared co-occurrence regularities. The goal
of the present experiment was to evaluate these predictions.
Present Experiment
The present experiment tested the Co-Occurrence account’s
prediction that associative links that can be learned from
direct co-occurrence develop early, and are gradually
supplemented by taxonomic links that can be learned from
shared co-occurrence. We therefore measured the
development of associative and taxonomic links between
familiar concepts from early childhood (4-year-old children)
to adulthood. To target associative and taxonomic links, we
measured the strength of semantic links between “associated”
concepts whose labels regularly co-occur in child language
input (MacWhinney, 2000), and “taxonomically related”
concepts similar in meaning ("About wordnet," 2010).
We designed our measurement of semantic links to fulfill
two important criteria. First, instead of merely measuring
whether a certain type of semantic link is present or absent in
a given age group, we acquired fine grained, sensitive
measures of the strength of semantic links. Such sensitive
measures are important for tracking the gradual,
developmental emergence of semantic links. Second, we
designed our measurement to primarily capture
developmental changes in semantic links, rather than in other
cognitive processes such as reasoning.
To fulfill these criteria, we used a Visual World paradigm.
This paradigm capitalizes on the fact that people tend to look
organization (McNeill, 1963). For the present research, the key point
is that direct co-occurrence can be directly experienced from
language input, whereas shared co-occurrence-based links can only
be derived by integrating across separate episodes of direct co-
at images that they perceive as related to language that they
hear. Therefore, we can measure the strength of semantic
links from the degree to which hearing a label for one concept
prompts looking at an image of another concept over time.
This measure is both fine-grained, and because it is based on
spontaneous looking behavior, should be relatively
uncontaminated by other cognitive processes.
In our Visual World paradigm, participants saw a pair of
unrelated Target pictures (e.g., bed and fish), and heard
either: (1) An Associate Prime for one of the Targets (e.g.,
pillow or water), (2) A Taxonomic Prime for one of the
Targets (e.g., chair or bird), or (3) An Unrelated prime that
was neither associated with nor taxonomically related to
either Target (e.g., stick). We measured the strength of
associative and taxonomic links based on the degree to which
participants looked more at Targets over time following
Associate or Taxonomic versus Unrelated Primes.
Importantly, our use of the same pairs of Target pictures in
all Prime conditions meant that looking differences across
Prime conditions can be attributed to semantic links between
Prime and Target concepts, rather than the salience, visual
properties, or subjective appeal Targets themselves.
Informed consent was obtained from parents/guardians of
child participants and from adult participants prior to
participation. The sample included 41 4-year-olds and 37
adults. Children were recruited from families, daycares, and
preschools and adults were recruited from the undergraduate
population at a public university in the same city.
The stimuli for this experiment were Sets of words
consisting of a Target, an Associate Prime, and a Taxonomic
Prime generated according to the following criteria.
Associate criteria. Associate Primes were selected as
words that reliably co-occur with Targets in corpora of child
speech input (CHILDES database; MacWhinney, 2000).
Using scripts developed in-lab, we measured the degree to
which word pairs co-occurred more frequently within a 7-
word window across 25 CHILDES corpora (O) than the
frequency with which they would be expected to co-occur by
chance, based on their respective frequencies (E). The larger
the difference between observed versus expected frequency,
the more reliably words in a pair co-occur. This ratio is
captured by the following “t.score” formula:
𝑡. 𝑠𝑐𝑜𝑟𝑒 = 𝑂 − 𝐸
Candidate Target-Associate pairs were pairs of nouns with
t-scores of > 2.5 (Baayen, Davidson, & Bates, 2008). In
addition, Associates could not meet the taxonomic criteria
described below.
Taxonomic criteria. Candidate Taxonomic Primes for
Targets were identified based on their membership in the
same taxonomic category (e.g., clothing, foods) and
similarity in meaning in WordNet (a database of word
definitions composed by lexicographers; 2010). In WordNet,
words are hierarchically organized such that more specific
words (e.g., dog) are subsumed within less specific words
(e.g., animal). We used Resnik similarity as a measure of
meaning similarity in WordNet, which is based on
identifying the most specific subsumer of a pair of words:
The more specific the subsumer, the higher the similarity. For
example, dog and cat are subsumed within carnivore,
whereas dog and rat are subsumed within mammal; because
carnivore is more specific than mammal, Resnik similarity is
higher between dog and cat versus dog and rat. Candidate
Taxonomic Primes had Resnik similarities to Targets of > 5,
and did not meet the Associate criterion.
Composition of Set Pairs. We used the Associate and
Taxonomic criteria to compose 22 Sets each consisting of a
Target, Associate, and a Taxonomic Prime. Importantly,
Targets were neither taxonomically related to Associate
Primes, nor associated with Taxonomic Primes. All words
also met a familiarity criterion of being produced by at least
55% of 36-month-old children (approximately one year
younger than children in our youngest sample). Words within
Sets were balanced for this criterion.
We organized these 22 Sets into 11 Set Pairs. Within Set
Pairs: (1) Targets were unrelated and equivalently familiar,
and (2) Primes for one Target were unrelated to the other
Target. To each Set Pair we added an Unrelated Prime that:
(1) Met the familiarity criterion, and (2) Met neither
Associate nor the Taxonomic criteria for both Targets.
Materials. All words in Set Pairs were recorded by a
female speaker using child-friendly speech. Targets were
presented as pictures subtending ~5.3 of visual angle.
This experiment used an EyeLink Portable Duo eye tracking
system with a sampling rate of 500Hz, and a button box that
participants used in a cover task (see Procedure).
Adults were tested in a quiet lab room, and children were
tested either in a quiet lab room, or at their preschool or
daycare. The procedure was similar for adults and children,
with the exception that children completed one block of trials,
and adults completed two blocks (i.e., repeated the same
block twice with randomized trial orders).
Figure 2: Sequence of events in experimental trials.
Following eye tracker calibration, the experiment consisted
of trials in which participants were shown the two Target
pictures from a Set Pair (e.g. apple and bottle), one presented
on the left side of the screen, on a yellow background, and
one on the right side of the screen, on a blue background. As
shown in Figure 2, the two Targets appeared alone for 500ms,
and then participants heard a word.
Participants first completed familiarization trials, then the
main experiment, which included a mix of cover task and
experimental trials. In familiarization and cover task trials,
participants heard “yellow” or “blue”, and clicked a button of
the same color on a button box to complete the trial. These
trials were designed to keep participants engaged in a task on
non-experimental trials.
Experimental trials were similar to familiarization/cover
task trials. However, instead of “yellow”/“blue”, participants
heard Primes from the Set Pairs. Participants were instructed
not to respond to these trials, which instead ended
automatically 2000ms following Prime onset.
Across trials, each pair of Target pictures (e.g., apple and
bottle) was presented with the five Primes from their Set Pair:
(1) The Associate Prime for one of the two Targets (e.g., tree
or baby), (2) The Taxonomic Prime for one of the two Targets
(e.g., grapes or bowl), or (3) The Unrelated Prime (e.g., door).
Thus, there was a total of 55 experimental trials within a
block: 22 Associate, 22 Taxonomic and 11 Unrelated. These
trials were mixed with 22 cover task trials (one “yellow” and
one “blue” trial for each Set Pair). The assignment of Target
pictures in each Set Pair to appear on the yellow background
on the left or blue background on the right was
counterbalanced across experimental and cover task trials.
This design ensured that looking on experimental trials was
not contaminated by response-related behavior.
To test the contributions of association and taxonomic
relatedness, the data from this experiment were used to
compare the time course of looking at Targets accompanied
by Associate or Taxonomic Primes versus Unrelated Primes
in each age group. To conduct this comparison, we first
generated outcome variables of interest.
Outcome Variables
Data from practice and filler trials were excluded from
analyses. The raw eye tracking data consisted of the position
of gaze on the screen sampled every 2ms within experimental
trials, which was identified as falling within an AOI for the
image on the left, an AOI for the image on the right, or neither
AOI. We removed data from the 500ms prior to Prime onset,
then divided the remaining two seconds into 100ms time bins.
We used these data to generate two outcome variables.
Target Dwell Time. We first calculated a Target Dwell
Time variable that captures the amount of time spent looking
at each Target in each time bin when accompanied by its
Associate, Taxonomic, or Unrelated Prime. We used this
variable to test whether looking dynamics for Targets
differed when accompanied by their Associate or Taxonomic
versus Unrelated Primes.
Difference from Unrelated. This variable captured the
degree to which looking in the Associate and Taxonomic
Prime conditions each deviated from the Unrelated Prime
condition. We calculated this value by subtracting the
Unrelated Target Dwell Time for a Target/time bin from both
the corresponding Target Dwell Time in the Associate
condition, and the in the Taxonomic condition. We used this
variable to test for differences between the effects of
Associate versus Taxonomic (relative to Unrelated) Primes.
Analysis of Looking Behavior
We followed the Growth Curve Analysis (GCA) approach
developed by Mirman and colleagues (Mirman, Dixon, &
Magnuson, 2008). GCA involves generating hierarchical
mixed effects models, starting with a “base” model with
temporal terms that captures how looking changes over time,
without considering variation across conditions, individuals,
or items. In the base model, the intercept captures the average
value of the looking outcome variable, a linear term captures
monotonic changes in the value of the outcome variable over
time, and a quadratic term captures the sharpness of the peak
in looking. Finally, cubic and quartic terms capture changes
in asymptotic tails of looking over time that are not typically
informative about the effects of experimental conditions.
To analyze the effects of experimental conditions, the base
model is supplemented with: Fixed effects of experimental
conditions and their interaction with the temporal terms,
random intercepts for participants and/or items, and random
slopes for effects of experimental conditions within
participants and/or items. Effects of experimental conditions
are interpreted from their interactions with temporal terms.
For example, an interaction between a fixed effect of
condition and the linear term reveals that condition influences
the monotonic increase or decrease in looking over time.
Target Dwell Time Analysis. We first tested whether the
temporal dynamics of looking at Targets differed when
accompanied by Associate or Taxonomic Primes in
comparison to when accompanied by Unrelated Primes.
Specifically, we generated separate models of Dwell Times
for Targets in each time bin for each age group that both
supplemented the base model with a fixed effect of Prime
condition (with Unrelated as the reference level to which
Associate and Taxonomic were compared). These models
Est. (SE)
Est. (SE)
9.599 (1.852)
6.389 (1.852)
31.351 (5.603)
13.229 (5.603)
-4.702 (5.057)
-9.130 (5.057)
8.986 (2.605)
6.768 (2.605)
21.462 (7.149)
14.384 (7.149)
-23.151 (5.097)
-18.836 (5.097)
additionally included random intercepts for participant and
item, and random slopes for the effect of Prime condition
within participants and within items.
Parameter estimates and their significance are reported in
Table 1. Both children and adults looked more overall at
Targets upon hearing an Associate or a Taxonomic versus an
Unrelated Prime (as shown by significant effects on the
Intercept). Associate and Taxonomic Primes also affected
changes in looking at a given Target over time, including the
rate at which looking at the Target increased (Linear term)
and/or the sharpness of the peak in Target looking time
(Quadratic term). Taken together, these results show that
concepts depicted by Targets were activated by both Co-
Occur and Taxonomic Primes in both adults and children.
Difference from Unrelated. This analysis assessed
differences in degree to which Associate and Taxonomic
Primes activated Targets, relative to Unrelated Primes.
Specifically, we generated separate models of Difference
from Unrelated values for children and adults that
supplemented the base model with a fixed effect of
Relatedness condition (Associate and Taxonomic), random
intercepts for participant and item, and random slopes for the
effect of Relatedness condition within participants and within
items. Figure 3 depicts the Difference from Unrelated data
and the corresponding fitted data from the models.
The parameter estimates and their significance levels are
reported in Table 2. In children, Associate Primes produced
grater rates of increased looking at Targets (relative to
Unrelated Primes) than Taxonomic Primes. In contrast, in
adults, no such differences were observed: Associate and
Taxonomic Primes affected looking at Targets relative to
Unrelated Primes to equivalent extents.
The present experiment revealed that, in young children,
associative links between concepts whose labels reliably co-
occur are initially stronger than taxonomic links between
concepts whose labels often share patterns of co-occurrence.
By adulthood however, associative and taxonomic links were
similar in strength. This trajectory is consistent with the Co-
Occurrence account prediction that associative links emerge
early, and are supplemented by taxonomic links with
development. These results therefore highlight how rich
semantic structure may emerge from simple but powerful
sensitivities to co-occurrence statistics.
However, the trajectory of semantic organization
development cannot be inferred from the present experiment
alone. To contextualize these findings, we next evaluate the
degree to which this developmental trajectory is consistent
with evidence from prior research on semantic development.
In this evaluation, we highlight how the present findings are
both consistent with, and expand upon much of the large body
of prior semantic development research.
Contribution of Co-Occurrence
Although a role for co-occurrence throughout semantic
organization development has been overlooked (or posited to
be transient) in the majority existing semantic development
accounts, the present evidence supporting this role is
consistent with many prior findings. Specifically, numerous
studies with children (e.g., Blaye, Bernard-Peyron, Paour, &
Bonthoux, 2006; Lucariello et al., 1992) and a handful of
studies with adults (e.g., Lin & Murphy, 2001) have observed
the presence of links in semantic organization that may be
learned from co-occurrence, such as schematic and thematic
relatedness. Moreover, in contrast with schematic and
thematic relatedness, which are constructs subjectively
defined by researchers, the present findings highlight co-
occurrence regularities as a measurable source of input in the
environment that may shape these semantic links.
Contribution of Taxonomic Relations
This experiment revealed an influence of taxonomic
relatedness that was initially weaker than the influence of co-
occurrence, but reached similar strength by adulthood.
Contextualizing this finding within prior research is
complicated by the fact that it has yielded conflicting
findings. One body of findings suggest that taxonomic
relations only gradually emerge starting from mid-to-late
childhood (e.g., age 6-7, Blaye et al., 2006; Lucariello et al.,
1992). In contrast, a similarly large body of findings suggests
that taxonomic relations are strong and robust starting early
in development (Gelman & Markman, 1986; Waxman &
Namy, 1997). In spite of these apparently contradictory
bodies of evidence, we suggest that the present findings can
be reconciled and shed new light on both.
Table 2: Difference from Unrelated GCA results.
Estimates are for the Associate versus the Taxonomic
condition. Non-significant estimates are in italics.
Associate vs Taxonomic
Est. (SE)
3.210 (1.970)
18.122 (5.938)
4.428 (5.094)
2.217 (2.493)
7.078 (7.541)
-4.315 (5.169)
Figure 3: Difference from Unrelated values in the
Associate and Taxonomic conditions in Children and
Adults, plotted with lines depicting the fitted values from
the models. Error bars show standard errors of the mean.
Gradual Taxonomic Development. Numerous studies
using a variety of behavioral paradigms have found that
taxonomic relations only begin to gradually contribute to
semantic organization with starting in mid-to-late childhood
(e.g., Blaye et al., 2006; Lucariello et al., 1992). The present
findings are consistent with evidence for the gradual
development of taxonomic relations, and suggest that
sensitive measures (such as those used in the present
experiment) can capture this gradual development starting
earlier in childhood.
Early Taxonomic Onset. Another large body of findings
suggests early, robust taxonomic organization. Many such
studies have assessed semantic organization using match-to-
sample paradigms, in which participants choose to match a
sample item (e.g., dog) with one of two other items (e.g.,
elephant and bone) that are related to the sample in different
ways. In some studies using variants of this paradigm (e.g.,
Bauer & Mandler, 1989; Gelman & Markman, 1986;
Waxman & Namy, 1997), young children chose taxonomic
matches throughout the study or under specific conditions.
Our findings also suggest the presence of taxonomic
relations in young children. Only the notion that these prior
findings indicate robust taxonomic knowledge starting in
early childhood conflicts with the present evidence that
taxonomic relations are initially weak. However, this
contradiction can be resolved by considering how additional
information that could support taxonomic choices was
available in prior studies showing “robust” taxonomic
relations in young children. For example, in some prior
studies, many target items are likely to have been visually
similar to (e.g., car and jeep, pot and skillet) and/or co-
occurring with (e.g., chair and table) their taxonomic
matches. Moreover, targets and taxonomic matches were
sometimes given either identical labels, which may act as
perceptual features that contribute to similarity in young
children (Sloutsky & Fisher, 2004), or co-occurring labels
(e.g., puppy and dog), such that taxonomic choices could be
based on co-occurrence. The availability of co-occurrence
and/or perceptual similarity in addition to taxonomic
relatedness also characterizes stimuli used in many studies of
semantic knowledge in infants (e.g., Willits, Wojcik,
Seidenberg, & Saffran, 2013).
Asynchronous Development of Associative and
Taxonomic Relations
The Co-Occurrence account predicts that associative links
emerge early because they can be formed from direct co-
occurrence regularities that can be directly gleaned from
language input. For example, hearing “I’d like a juicy apple
can immediately contribute to a semantic link between
“juicy” and “apple”. By the same token, the Co-occurrence
account predicts that taxonomic links emerge later because
they may rely on shared co-occurrence regularities that can
only be derived by integrating across separate episodes of
direct co-occurrence. However, the Co-Occurrence account
does not currently specify a precise reason for why
associative links based on direct co-occurrence may form
earlier in development than taxonomic links based on shared
co-occurrence. Instead, the present evidence for the Co-
Occurrence account highlights explanations for this
developmental asynchrony to be explored in future research.
One possibility is that taxonomic links develop more
slowly simply because they require more language input:
Hearing juicy and apple directly co-occurring can
immediately contribute to an associative link, whereas the
learner must separately hear juicy with both apple and
pear to form a shared co-occurrence-based taxonomic link.
Alternatively, abilities to form links between inputs based on
direct and shared co-occurrence statistics may themselves
develop asynchronously. This possibility is supported by the
contrast between extensive statistical learning evidence that
even infants can form links between inputs that directly co-
occur (Fiser & Aslin, 2002; Saffran et al., 1996; Saffran,
Johnson, Aslin, & Newport, 1999), and a handful of evidence
for the more gradual development of abilities to form links
based on shared co-occurrence (Bauer & San Souci, 2010;
Schlichting et al., 2017). Disentangling these possibilities can
shed light on the underlying processes that drive
developmental changes in semantic organization.
Organized semantic knowledge plays a fundamental role in
many facets of human intelligence. The present experiment
provides evidence supporting the possibility that this
organization emerges in part from the operation of simple but
powerful learning mechanisms that form semantic links from
statistical regularities in language.
Princeton University. (2010). About wordnet.
Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008).
Mixed-effects modeling with crossed random effects for
subjects and items. Journal of Memory and Language, 59,
Bauer, P. J., & Larkina, M. (2017). Realizing relevance: The
influence of domainspecific information on generation of
new knowledge through integration in 4to 8yearold
children. Child Development, 88, 247-262.
Bauer, P. J., & Mandler, J. M. (1989). Taxonomies and triads:
Conceptual organization in one-to two-year-olds.
Cognitive Psychology, 21, 156-184.
Bauer, P. J., & San Souci, P. (2010). Going beyond the facts:
Young children extend knowledge by integrating episodes.
Journal of Experimental Child Psychology, 107, 452-465.
Blaye, A., Bernard-Peyron, V., Paour, J.-L., & Bonthoux, F.
(2006). Categorical flexibility in children: Distinguishing
response flexibility from conceptual flexibility. European
Journal of Developmental Psychology, 3, 163-188.
Fiser, J., & Aslin, R. N. (2002). Statistical learning of new
visual feature combinations by infants. Proceedings of the
National Academy of Sciences, 99, 15822-15826.
Gelman, S. A., & Markman, E. M. (1986). Categories and
induction in young children. Cognition, 23, 183-209.
Hall, G., Mitchell, C., Graham, S., & Lavis, Y. (2003).
Acquired equivalence and distinctiveness in human
discrimination learning: Evidence for associative
mediation. Journal of Experimental Psychology: General,
132, 266.
Hofmann, M. J., Biemann, C., Westbury, C., Murusidze, M.,
Conrad, M., & Jacobs, A. M. (2018). Simple cooccurrence
statistics reproducibly predict association ratings.
Cognitive Science, 42, 2287-2312.
Honey, R. C., & Hall, G. (1989). Acquired equivalence and
distinctiveness of cues. Journal of Experimental
Psychology: Animal Behavior Processes, 15, 338.
Inhelder, B., & Piaget, J. (1964). The early growth of logic in
the child. New York: Norton.
Jones, M. N., Willits, J., & Dennis, S. (2015). Models of
semantic memory. In J. Busemeyer & J. Townsend (Eds.),
Oxford handbook of mathematical and computational
psychology (pp. 232-254). New York, NY: Oxford
University Press.
Landauer, T. K., & Dumais, S. T. (1997). A solution to plato's
problem: The latent semantic analysis theory of
acquisition, induction, and representation of knowledge.
Psychological Review, 104, 211.
Lin, E. L., & Murphy, G. L. (2001). Thematic relations in
adults' concepts. Journal of Experimental Psychology:
General, 130, 3-28.
Lucariello, J., Kyratzis, A., & Nelson, K. (1992). Taxonomic
knowledge: What kind and when? Child Development, 63,
MacWhinney, B. (2000). The childes project: The database
(Vol. 2): Psychology Press.
Matlen, B. J., Fisher, A. V., & Godwin, K. E. (2015). The
influence of label co-occurrence and semantic similarity on
children’s inductive generalization. Frontiers in
Psychology, 6, 1146.
McNeill, D. (1963). The origin of associations within the
same grammatical class. Journal of Verbal Learning and
Verbal Behavior, 2, 250-262.
Mirman, D., Dixon, J. A., & Magnuson, J. S. (2008).
Statistical and computational models of the visual world
paradigm: Growth curves and individual differences.
Journal of Memory and Language, 59, 475-494.
Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996).
Statistical learning by 8-month-old infants. Science, 274,
Saffran, J. R., Johnson, E. K., Aslin, R. N., & Newport, E. L.
(1999). Statistical learning of tone sequences by human
infants and adults. Cognition, 70, 27-52.
Sahlgren, M. (2008). The distributional hypothesis. Italian
Journal of Disability Studies, 20, 33-53.
Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne,
N. B., & Botvinick, M. M. (2013). Neural representations
of events arise from temporal community structure. Nature
Neuroscience, 16, 486-492.
Schapiro, A. C., Turk-Browne, N. B., Botvinick, M. M., &
Norman, K. A. (2017). Complementary learning systems
within the hippocampus: A neural network modelling
approach to reconciling episodic memory with statistical
learning. Philosophical Transactions of the Royal Society
B: Biological Sciences, 372, 20160049.
Schlichting, M. L., Guarino, K. F., Schapiro, A. C., Turk-
Browne, N. B., & Preston, A. R. (2017). Hippocampal
structure predicts statistical learning and associative
inference abilities during development. Journal of
Cognitive Neuroscience, 29, 37-51.
Sloutsky, V. M. (2010). From perceptual categories to
concepts: What develops? Cognitive Science, 34, 1244-
Sloutsky, V. M., & Fisher, A. V. (2004). Induction and
categorization in young children: A similarity-based
model. Journal of Experimental Psychology: General, 133,
Sloutsky, V. M., Yim, H., Yao, X., & Dennis, S. (2017). An
associative account of the development of word learning.
Cognitive Psychology, 97, 1-30.
Spence, D. P., & Owens, K. C. (1990). Lexical co-occurrence
and association strength. Journal of Psycholinguistic
Research, 19, 317-330.
Waxman, S. R., & Namy, L. L. (1997). Challenging the
notion of a thematic preference in young children.
Developmental Psychology, 33, 555-567.
Willits, J. A., Wojcik, E. H., Seidenberg, M. S., & Saffran, J.
R. (2013). Toddlers activate lexical semantic knowledge in
the absence of visual referents: Evidence from auditory
priming. Infancy, 18, 1053-1075.
Wojcik, E. H., & Saffran, J. R. (2015). Toddlers encode
similarities among novel words from meaningful
sentences. Cognition, 138, 10-20.
Zeithamova, D., Dominick, A. L., & Preston, A. R. (2012).
Hippocampal and ventral medial prefrontal activation
during retrieval-mediated learning supports novel
inference. Neuron, 75, 168-179.
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
What determines human ratings of association? We planned this paper as a test for association strength (AS) that is derived from the log likelihood that two words co‐occur significantly more often together in sentences than is expected from their single word frequencies. We also investigated the moderately correlated interactions of word frequency, emotional valence, arousal, and imageability of both words (r's ≤ .3). In three studies, linear mixed effects models revealed that AS and valence reproducibly account for variance in the human ratings. To understand further correlated predictors, we conducted a hierarchical cluster analysis and examined the predictors of four clusters in competitive analyses: Only AS and word2vec skip‐gram cosine distances reproducibly accounted for variance in all three studies. The other predictors of the first cluster (number of common associates, (positive) point‐wise mutual information, and word2vec CBOW cosine) did not reproducibly explain further variance. The same was true for the second cluster (word frequency and arousal); the third cluster (emotional valence and imageability); and the fourth cluster (consisting of joint frequency only). Finally, we discuss emotional valence as an important dimension of semantic space. Our results suggest that a simple definition of syntagmatic word contiguity (AS) and a paradigmatic measure of semantic similarity (skip‐gram cosine) provide the most general performance‐independent explanation of association ratings.
Full-text available
A growing literature suggests that the hippocampus is critical for the rapid extraction of regularities from the environment. Although this fits with the known role of the hippocampus in rapid learning, it seems at odds with the idea that the hippocampus specializes in memorizing individual episodes. In particular, the Complementary Learning Systems theory argues that there is a computational trade-off between learning the specifics of individual experiences and regularities that hold across those experiences. We asked whether it is possible for the hippocampus to handle both statistical learning and memorization of individual episodes. We exposed a neural network model that instantiates known properties of hippocampal projections and subfields to sequences of items with temporal regularities. We found that the monosynaptic pathway—the pathway connecting entorhinal cortex directly to region CA1—was able to support statistical learning, while the trisynaptic pathway—connecting entorhinal cortex to CA1 through dentate gyrus and CA3—learned individual episodes, with apparent representations of regularities resulting from associative reactivation through recurrence. Thus, in paradigms involving rapid learning, the computational trade-off between learning episodes and regularities may be handled by separate anatomical pathways within the hippocampus itself. This article is part of the themed issue ‘New frontiers for statistical learning in the cognitive sciences’.
Full-text available
Semantically-similar labels that co-occur in child-directed speech (e.g., bunny-rabbit) are more likely to promote inductive generalization in preschoolers than non-co-occurring labels (e.g., lamb-sheep). However, it remains unclear whether this effect stems from co-occurrence or other factors, and how co-occurrence contributes to generalization. To address these issues, preschoolers were exposed to a stream of semantically-similar labels that don't co-occur in natural language, but were arranged to co-occur in the experimental setting. In Experiment 1, children exposed to the co-occurring stream were more likely to make category-consistent inferences than children in two control conditions. Experiment 2 replicated this effect and provided evidence that co-occurrence training influenced generalization only when the trained labels were categorically-similar. These findings suggest that both co-occurrence information and semantic representations contribute to preschool-age children's inductive generalization. The findings are discussed in relation to the developmental accounts of inductive generalization.
Word learning is a notoriously difficult induction problem because meaning is underdetermined by positive examples. How do children solve this problem? Some have argued that word learning is achieved by means of inference: young word learners rely on a number of assumptions that reduce the overall hypothesis space by favoring some meanings over others. However, these approaches have difficulty explaining how words are learned from conversations or text, without pointing or explicit instruction. In this research, we propose an associative mechanism that can account for such learning. In a series of experiments, 4-year-olds and adults were presented with sets of words that included a single nonsense word (e.g. dax). Some lists were taxonomic (i.,e., all items were members of a given category), some were associative (i.e., all items were associates of a given category, but not members), and some were mixed. Participants were asked to indicate whether the nonsense word was an animal or an artifact. Adults exhibited evidence of learning when lists consisted of either associatively or taxonomically related items. In contrast, children exhibited evidence of word learning only when lists consisted of associatively related items. These results present challenges to several extant models of word learning, and a new model based on the distinction between syntagmatic and paradigmatic associations is proposed.
One of the primary functions of natural kind terms (e.g., tiger, gold) is to support inductive inferences. People expect members of such categories to share important, unforeseen properties, such as internal organs and genetic structure. Moreover, inductions can be made without perceptual support: even when an object does not look much like other members of its category, and even when a property is unobservable. The present work addresses how expectations about natural kinds originate. Young children, with their usual reliance on perceptual appearances and only rudimentary scientific knowledge, might not induce new information within natural kind categories. To test this possibility, category membership was pitted against perceptual similarity in an induction task. For example, children had to decide whether a shark is more likely to breathe as a tropical fish does because both are fish, or as a dolphin does because they look alike. By at least age 4, children can use categories to support inductive inferences even when category membership conflicts with appearances. Moreover, these young children have partially separated out properties that support induction within a category (e.g., means of breathing) from those that are in fact determined by perceptual appearances (such as weight). Since we examined only natural kind categories, we do not know to what extent children have differentiated natural kinds from other sorts of categories. Children may start out assuming that categories named by language have the structure of natural kinds and with development refine these expectations.
Despite the importance of learning and remembering across the lifespan, little is known about how the episodic memory system develops to support the extraction of associative structure from the environment. Here, we relate individual differences in volumes along the hippocampal long axis to performance on statistical learning and associative inference tasks-both of which require encoding associations that span multiple episodes-in a developmental sample ranging from ages 6 to 30 years. Relating age to volume, we found dissociable patterns across the hippocampal long axis, with opposite nonlinear volume changes in the head and body. These structural differences were paralleled by performance gains across the age range on both tasks, suggesting improvements in the cross-episode binding ability from childhood to adulthood. Controlling for age, we also found that smaller hippocampal heads were associated with superior behavioral performance on both tasks, consistent with this region's hypothesized role in forming generalized codes spanning events. Collectively, these results highlight the importance of examining hippocampal development as a function of position along the hippocampal axis and suggest that the hippocampal head is particularly important in encoding associative structure across development.
In accumulating knowledge, direct modes of learning are complemented by productive processes, including self-generation based on integration of separate episodes. Effects of the number of potentially relevant episodes on integration were examined in 4- to 8-year-olds (N = 121; racially/ethnically heterogeneous sample, English speakers, from large metropolitan area). Information was presented along with unrelated or related episodes; the latter challenged children to identify the relevant subset of episodes for integration. In Experiment 1, 4- and 6-year-olds integrated in the unrelated context. Six-year-olds also succeeded in the related context in forced-choice testing. In Experiment 2, 8-year-olds succeeded in open-ended and forced-choice testing. Results illustrate a developmental progression in productive extension of knowledge due in part to age-related increases in identification of relevant information.
Language learners rapidly acquire extensive semantic knowledge, but the development of this knowledge is difficult to study, in part because it is difficult to assess young children's lexical semantic representations. In our studies, we solved this problem by investigating lexical semantic knowledge in 24-month-olds using the Head-turn Preference Procedure. In Experiment 1, looking times to a repeating spoken word stimulus (e.g., kitty-kitty-kitty) were shorter for trials preceded by a semantically related word (e.g., dog-dog-dog) than trials preceded by an unrelated word (e.g., juice-juice-juice). Experiment 2 yielded similar results using a method in which pairs of words were presented on the same trial. The studies provide evidence that young children activate of lexical semantic knowledge, and critically, that they do so in the absence of visual referents or sentence contexts. Auditory lexical priming is a promising technique for studying the development and structure of semantic knowledge in young children.
A tradition of research on conceptual development has suggested a shift from a thematic to a taxonomic basis of organization near the end of the preschool years. Recent research has demonstrated an increase in taxonomic responding in preschool-age subjects with provision of novel labels for to-be-classified targets. The effect is interpreted as the result of a cognitive constraint on the possible meanings of new words. In the present series of four experiments, 16- to 31-month-olds responded to a forced-choice object-triad task. Across the age range, subjects responded taxonomically at well above chance levels. Subjects as young as 16 months responded taxonomically 72% of the time on basic-level triads; subjects as young as 19 months responded taxonomically 87% of the time on both basic- and superordinate-level triads. The already high rate of taxonomic responding was not affected by novel labels. In control experiments, subjects demonstrated familiarity with the thematic associates depicted in the test triads. In all four experiments, nonverbal reinforcement was used to unambiguously convey task instructions and to remind subjects of them on each trial. We suggest that verbal labels may accomplish the same end for older preschoolers. Thus, the effect of labeling, rather than being language specific, may be a more general effect of reminding subjects of the requirements of the task.