PreprintPDF Available

Abstract and Figures

Our knowledge about the world is represented not merely as a collection of concepts, but as an organized lexico-semantic network in which concepts can be linked by relations, such as “taxonomic” relations between members of the same stable category (e.g., cat and sheep), or association between entities that occur together or in the same context (e.g., sock and foot). To date, accounts of the origins of semantic organization have largely overlooked how sensitivity to statistical regularities ubiquitous in the environment may play a powerful role in shaping semantic development. The goal of the present research was to investigate how associations in the form of statistical regularities with which labels for concepts co-occur in language (e.g., sock and foot) and taxonomic relatedness (e.g., sock and pajamas) shape semantic organization of 4-5-year-olds and adults. To examine these aspects of semantic organization across development, we conducted three experiments examining effects of co-occurrence and taxonomic relatedness on cued recall (Experiment 1), word-picture matching (Experiment 2), and looking dynamics in a Visual World paradigm (Experiment 3). Taken together, the results of the three experiments provide evidence that co-occurrence-based links between concepts manifest in semantic organization from early childhood onward, and are increasingly supplemented by taxonomic links. We discuss these findings in relation to theories of semantic development.
Content may be subject to copyright.
Statistical Regularities Shape Semantic Organization throughout Development
Layla Unger
Olivera Savic
Vladimir M. Sloutsky
Department of Psychology, Ohio State University, Columbus OH
Acknowledgements
This work was supported by National Institutes of Health Grants R01HD078545 and
P01HD080679 to Vladimir Sloutsky. We additionally thank Taylor Swenski for her
contributions to this research.
Data Availability Statement
Data and scripts for processing and analyzing data have been made available through
the Open Science Framework at https://osf.io/ecm9u/.
Please address correspondence to:
Layla Unger
Ohio State University, Department of Psychology
1835 Neil Avenue, PS 268
Columbus, Ohio 43210
Phone: (614) 688-1235
Email: unger.114@osu.edu
Statistical Regularities Shape Semantic Development 1
Abstract
Our knowledge about the world is represented not merely as a collection of concepts, but
as an organized lexico-semantic network in which concepts can be linked by relations,
such as “taxonomic” relations between members of the same stable category (e.g., cat
and sheep), or association between entities that occur together or in the same context
(e.g., sock and foot). To date, accounts of the origins of semantic organization have
largely overlooked how sensitivity to statistical regularities ubiquitous in the environment
may play a powerful role in shaping semantic development. The goal of the present
research was to investigate how associations in the form of statistical regularities with
which labels for concepts co-occur in language (e.g., sock and foot) and taxonomic
relatedness (e.g., sock and pajamas) shape semantic organization of 4-5-year-olds and
adults. To examine these aspects of semantic organization across development, we
conducted three experiments examining effects of co-occurrence and taxonomic
relatedness on cued recall (Experiment 1), word-picture matching (Experiment 2), and
looking dynamics in a Visual World paradigm (Experiment 3). Taken together, the results
of the three experiments provide evidence that co-occurrence-based links between
concepts manifest in semantic organization from early childhood onward, and are
increasingly supplemented by taxonomic links. We discuss these findings in relation to
theories of semantic development.
Keywords: semantic memory; semantic development; semantic organization; knowledge
organization; statistical regularities, cognitive development
Statistical Regularities Shape Semantic Development 2
Statistical Regularities Shape Semantic Organization throughout Development
Our knowledge about the world is fundamental to many cognitive feats we accomplish
on an everyday basis, including applying what we know to new situations (e.g., expecting
new home appliances to be powered by electricity), retrieving previously acquired
knowledge from memory, and incorporating new information into our existing body of
knowledge (Bower, Clark, Lesgold, & Winzenz, 1969; Heit, 2000; Jimura, Hirose, Wada
et al., 2016). These feats are possible because our knowledge about the world is not a
collection of isolated facts, but rather an interconnected lexico-semantic network of
related concepts (Cree & Armstrong, 2012; Jones, Willits, & Dennis, 2015; McClelland &
Rogers, 2003). For example, our knowledge about dogs is often connected to our
knowledge of other similar animals (such as cats), as well as to our knowledge about the
items with which dogs are associated in the environment, such as leashes, human
owners, and doghouses.
Although the fact that our concepts are organized is hardly controversial (Howard &
Howard, 1977; McClelland & Rogers, 2003; Ross & Murphy, 1999; Storm, 1980), the
processes that drive the development of semantic organization are a topic of considerable
debate. To date, this debate has primarily focused on how connections between concepts
from the same stable, “taxonomic” category (e.g., animals, foods) are formed, in spite of
the fact that taxonomic relatedness may be difficult to observe because members of the
same taxonomic category do not necessarily look similar, or occur together
1
.
1
Here and in most prior theoretical accounts and research into semantic organization, the challenge is to
explain the development of semantic links between different concepts (e.g., between dog and other
animals), not just the emergence of basic-level categories (e.g., dog), for which many cues (e.g., shared
labels and visual similarity) are readily available.
Statistical Regularities Shape Semantic Development 3
Some proposals suggest that the development of knowledge organization starts with
easy to observe relations, which then both bootstrap and are overwritten by knowledge
of taxonomic relations (Inhelder & Piaget, 1964; Lucariello, Kyratzis, & Nelson, 1992).
Alternately, other proposals suggest that we are endowed with early-emerging biases
towards learning taxonomic relations, such as a belief that items that are referred to by
the same label (e.g., “animal”) belong to the same taxonomic category (Fulkerson &
Waxman, 2007; Gelman & Coley, 1990; Gelman & Markman, 1986).
The goal of the current research is to investigate yet another possibility: That easy to
observe relations specifically, co-occurrence play a fundamental role in knowledge
organization from early in development through adulthood. Specifically, co-occurrence
may directly foster the formation of associations between corresponding concepts, thus
linking items such as spaghetti, fork, plate and napkin. Moreover, co-occurrence may also
indirectly foster the formation of links between concepts that share patterns of co-
occurrence, which are often taxonomically related, such as spaghetti and pie (which both
co-occur with fork, plate and napkin) (see Jones et al., 2015, for a review of mechanistic
models of forming relations from co-occurrence regularities). Additionally, because co-
occurrence regularities can be experienced directly, links between concepts based on
these regularities may manifest in semantic organization beginning early in development.
In contrast, shared patterns of co-occurrence must be integrated across separate
instances of co-occurrence. Therefore, learning taxonomic relations from shared patterns
of co-occurrence may require more time, such that taxonomic relations may emerge more
gradually in the course of development (Sloutsky, Yim, Yao, & Dennis, 2017). Importantly,
Statistical Regularities Shape Semantic Development 4
according to this view, taxonomic relations supplement rather than replace co-occurrence
relations.
In what follows, we first review extant theoretical accounts that have focused on
explaining the emergence of taxonomic relations in knowledge organization. We then
highlight key findings indicative of a role for co-occurrence that these accounts fail to
capture, and an alternate perspective in which co-occurrence contributes substantially to
semantic development. Finally, we present three experiments designed to examine the
presence of co-occurrence-based links and taxonomic links in semantic organization
across development.
Traditional Accounts of the Emergence of Semantic Organization
Most extant accounts of the development of semantic organization have focused on
how semantic knowledge becomes organized according to the membership of concepts
in stable, taxonomic categories, such as foods. According to some of these accounts,
referred to here as restructuring accounts, taxonomic relations are the endpoint of
maturational/learning processes that unfold across development. These perspectives fall
within a broader class of cognitive development accounts in which knowledge or abilities
present early in development are supplanted by their successors (e.g., Carey, 1985).
Critical to these accounts is the idea that the order in which concepts and their
relations are acquired is dictated by how observable they are. For example, it is easy to
observe that dogs reliably co-occur with leashes or bones. In contrast, the membership
of separate concepts in the same taxonomic category is often more difficult (if not
impossible) to observe: For example, animals can be quite different from one another,
and they do not necessarily appear together or in the same environment or context.
Statistical Regularities Shape Semantic Development 5
Restructuring accounts propose that early organization is shaped by information readily
available in the environment, and that over the course of development, taxonomic
knowledge overwrites (or replaces) this form of (more rudimentary) organization.
An early restructuring account was proposed by Inhelder and Piaget (1964).
According to this account, the transition towards taxonomic organization is driven by
experiences that highlight the inadequacy of earlier modes of organization (although the
mechanisms by which this transition occurs were not clearly explicated).
Another, more clearly specified restructuring account is Nelson and Lucariello’s (e.g.,
Lucariello et al., 1992) slot-filler account. This account highlights the fact that both
language and experience in the world contain regularities in which members of the same
taxonomic category often play the same role in the same context. For example, some
members of the taxonomic category of foods, such as pancakes, eggs, and bacon,
reliably play the role of being eaten in a breakfast context. According to the slot-filler
account, young children are sensitive to these regularities, such that semantic knowledge
first becomes organized into contextually-constrained taxonomic groups. With
development, contextually-independent taxonomic organization emerges as children
come to abstract across these constrained taxonomic groups and recognize when entities
play the same role in different contexts, such as foods being eaten in breakfast, lunch,
and dinner contexts.
According to another set of accounts, referred to here as taxonomic bias accounts,
taxonomic relations predominate semantic organization from early in development. This
type of account acknowledges that taxonomic relatedness is not directly apparent from
environmental input, but posits that it is nonetheless apprehended early in development
Statistical Regularities Shape Semantic Development 6
due to early-emerging (possibly innate) biases towards learning which entities are
members of the same taxonomic category. For example, members of the same basic-
level taxonomic category are often referred to by the same label, such as “bird”.
Accordingly, taxonomic bias accounts propose that we are endowed with early-emerging
beliefs that entities in the world belong to taxonomic categories, and that entities that are
referred to using the same label belong to the same category (Fulkerson & Waxman,
2007; Gelman & Coley, 1990; Gelman & Markman, 1986). Therefore, in these accounts,
learning that starts early in development consists of using these labels to anchor the
organization of semantic knowledge into basic-level taxonomic categories. However, it is
less clear how these biases support the formation of relations between these categories
that are crucial to semantic organization, such as relations between birds and other
animal categories. In addition, a role for other types of input, such as the regularity with
which entities co-occur, is not specified in these accounts.
A final type of account reviewed here, which we refer to as featural learning, posits
that the development of semantic organization is driven by detecting clusters of reliably
correlated features that are often associated with taxonomic categories (Rosch, 1975,
1978). For example, membership in the category of birds is associated with possessing
wings, feathers, and a beak. Featural learning accounts propose learning mechanisms
that are sensitive to these correlations, and that the operation of such mechanisms yields
taxonomic organization (Kemp & Tenenbaum, 2008; McClelland & Rogers, 2003). In
contrast with taxonomic bias accounts, featural learning accounts argue in favor of the
gradual emergence of taxonomic organization over the course of development. However,
like taxonomic bias accounts, featural learning accounts do not specify any role in
Statistical Regularities Shape Semantic Development 7
semantic organization for spatial or temporal co-occurrences between objects or the
words that denote them.
A Potential Role for Statistical Co-Occurrence Regularities
Of the influential accounts reviewed in the previous section, only some restructuring
accounts posit any role for co-occurrence regularities in semantic organization. However,
even in these accounts, these regularities play a role only early in development, and are
subsequently overwritten. At the same time, several findings highlight the potential
importance of co-occurrence regularities throughout development.
First, evidence from numerous studies suggests that sensitivity to co-occurrence
regularities (including the co-occurrence of words and objects) is apparent even early in
development (Bulf, Johnson, & Valenza, 2011; Samuelson & Smith, 1999; Wojcik &
Saffran, 2015). Moreover, numerous findings attest to effects on children’s reasoning of
semantic relations that may be formed based on co-occurrence, including schematic
relations between entities that occur in the same context (e.g., cow and barn) and
thematic relations between entities that play complementary roles (e.g., nail and hammer)
(Blaye, Bernard-Peyron, Paour, & Bonthoux, 2006; Fenson, Vella, & Kennedy, 1989;
Smiley & Brown, 1979; Tversky, 1985; Walsh, Richardson, & Faulkner, 1993). These
findings suggest that accounts of semantic development that do not posit any role for co-
occurrence between objects or their labels, such as the taxonomic bias and featural
learning accounts, are at best incomplete.
Second, evidence from a handful of studies suggests that semantic relations that may
be derived from co-occurrence continue to manifest in semantic organization into
adulthood. For example, in a series of ten experiments, Lin and Murphy (2001) observed
Statistical Regularities Shape Semantic Development 8
that relations between entities that adult raters judged as associated in scenes or events
(which likely co-occur in the environment) had a pervasive influence on adults’
categorization and reasoning that was frequently greater than the influence of taxonomic
relations (see also Ross & Murphy, 1999). This evidence is inconsistent with restructuring
accounts, in which an influence of co-occurrence early in development is eventually
overwritten. More broadly, the proposal inherent in restructuring accounts that later-
emerging knowledge and abilities replace those that emerge earlier in development is
also inconsistent with evidence that adults revert to childlike patterns of semantic
reasoning under cognitive load (Goldberg & Thompson-Schill, 2009).
Finally, the potential contributions of co-occurrence regularities are highlighted by a
mechanistic account and corroborating behavioral evidence presented by Sloutsky et al.
(2017). This account was inspired by computational modeling evidence that everyday
language input, including input to children (Asr, Willits, & Jones, 2016; Frermann &
Lapata, 2015; Huebner & Willits, 2018), is rich in statistical co-occurrence regularities that
capture links between concepts in semantic organization (see Jones et al., 2015 for a
review). First, regularities with which words directly co-occur, such as fork and spaghetti,
link concepts that are reliably associated in semantic knowledge (Hofmann, Biemann,
Westbury et al., 2018; Spence & Owens, 1990). Moreover, regularities with which words
share each other’s patterns of co-occurrence with other words, such as spaghetti and pie
(both co-occur with fork), link members of the same taxonomic category. Therefore,
according to Sloutsky et al.’s (2017) account, exposure to co-occurrence regularities in
language fosters both the learning of associations between concepts whose labels
directly co-occur, and between taxonomically related concepts whose labels share
Statistical Regularities Shape Semantic Development 9
patterns of co-occurrence. However, whereas direct co-occurrence can be gleaned
straight from the input and therefore rapidly foster links between concepts, shared
patterns of co-occurrence should foster links between members of the same taxonomic
category more slowly because they can only be derived by integrating across multiple
instances of direct co-occurrence. Moreover, a similar process may unfold for co-
occurrence patterns between the entities that concepts represent, given evidence that co-
occurrence patterns between words in language are closely mirrored by co-occurrence
patterns between objects in everyday visual scenes (Sadeghi, McClelland, & Hoffman,
2015). This account predicts both that: (1) Concepts should be linked based on direct co-
occurrence starting early and continuing throughout development, and (2) The linking of
taxonomically related concepts should gradually supplement co-occurrence links.
Initial support for this account comes from a series of word learning and lexical
extension experiments (Sloutsky et al., 2017) in which both children and adults had to
infer a meaning for a novel word embedded in a list of familiar words. When the novel
word appeared in a list of words that are associated (and therefore likely to co-occur) with
the word “animal”, such as “furry” and “zoo”, both children and adults inferred that the
novel word meant “animal”. In contrast, when the novel word appeared in a list of
members of the taxonomic category of animals such as “lion” and “bunny”, only adults
inferred this meaning. Moreover, this account is consistent with and can therefore
potentially help explain prior findings suggesting that taxonomic relations emerge later in
development than earlier-emerging relations such as associative links (e.g., Bjorklund &
Jacobs, 1985; Blaye et al., 2006; Fenson et al., 1989)
Statistical Regularities Shape Semantic Development 10
Together, these prior findings suggest that whereas direct co-occurrence-based links
may manifest in semantic organization throughout development and into adulthood,
taxonomic organization may gradually supplement these links because they are derived
(at least in part) from them. However, in addition to being overlooked in traditional
theoretical accounts of the development of semantic organization, this possibility has
received only limited empirical investigation to date. Furthermore, even when
investigated, actual co-occurrence in the environment has rarely been assessed. Instead,
researchers have investigated semantic relations between items that either are: (1)
Judged by the researchers themselves to co-occur, (2) Judged by adult raters to co-occur,
or (3) Produced by participants in free association tasks.
Although this approach has yielded evidence that is informative about the
development of semantic organization, researchers’ or raters’ judgments and free
associations do not directly estimate co-occurrence in environmental input, because
these judgements are themselves consequences of learning. This approach therefore
does not capture the nature of the input information that contributes to such learning
(Hofmann et al., 2018). For example, with respect to researcher and rater judgments,
there is evidence that judgments of co-occurrence are contaminated by taxonomic
relatedness (and vice versa; Wisniewski & Bassok, 1999). Similarly, the nature of the
relations that link words produced in free association tasks must be subjectively inferred
and can potentially be taxonomic, derived from co-occurrence, or some other type of link
such as a part-whole relation. A more direct estimate of co-occurrence regularities from
actual input may therefore provide a more accurate estimate of the role of co-occurrence
in semantic organization throughout development.
Statistical Regularities Shape Semantic Development 11
Current Study
The overall goal of the current study was to investigate the presence of co-occurrence
and taxonomic links in lexico-semantic organization across development, from early
childhood to adulthood. This investigation was designed to arbitrate between competing
theoretical accounts of the development of knowledge organization. Specifically,
restructuring accounts predict that co-occurrence should contribute to knowledge
organization in childhood, but be replaced by taxonomic relations in adulthood. Neither
taxonomic bias nor featural learning accounts make any predictions about the
contributions of co-occurrence. However, whereas the former predict that taxonomic
relations should contribute from childhood through to adulthood, the latter predict that the
contributions of taxonomic relations should substantially increase with age.
A different developmental pattern is predicted by recent proposals that highlight a key
role throughout development for co-occurrence in which it both directly fosters relations
between concepts, and indirectly fosters relations between concepts that share patterns
of co-occurrence and are often taxonomically related (e.g., Sloutsky et al., 2017).
Specifically, such proposals predict that relations between concepts whose labels or
referents regularly co-occur should be evident in the semantic organization of both
children and adults, and be increasingly supplemented by taxonomic relatedness over
the course of development.
We accomplished this investigation by measuring the degree to which familiar
concepts were related in young children (4- to 5-year-olds) and adults’ semantic
knowledge when either the concepts’ labels reliably co-occur in linguistic input, or when
they are members of the same taxonomic category. To target actual experienced co-
Statistical Regularities Shape Semantic Development 12
occurrence, we identified pairs of items based on the regularity with which the words for
a variety of concepts familiar to young children (e.g., cat, table) co-occurred more reliably
with each other than with other words in corpora of child-directed speech. To facilitate the
comparison between children and adults, we used paradigms developed to assess
semantic relatedness implicitly, without the requirement for engaging in and articulating
reasoning about relatedness, which adults may accomplish more easily. This approach
contrasts with previous studies that have used more explicit reasoning or generalization
tasks in an attempt to assess semantic organization across development (e.g., Sloutsky,
et al., 2017), in which developmental changes may in part be due to improvements in
abilities such as reasoning.
Paradigms. To obtain a generalizable representation of lexico-semantic organization
across development, the three reported experiments used three different paradigms
yielding implicit measures of semantic organization. The paradigms all share the same
underlying logic: When two concepts are linked, one should automatically activate the
other. However, the paradigms measure this automatic activation in different ways.
In Experiment 1, we used a Cued Recall paradigm to measure the effects of co-
occurrence and taxonomic relatedness on memory retrieval. The logic of this paradigm
was that links between a pair of concepts should facilitate the accuracy with which the
words for these concepts are recalled. Co-occurrence and taxonomic links were therefore
measured based on the degree to which they facilitated the recall of word pairs, in
comparison to unrelated pairs (e.g., Blewitt & Toppino, 1991).
In Experiment 2, we used a Match Verification paradigm in which participants
identified whether a word and a subsequent picture denoted the same item (e.g., the word
Statistical Regularities Shape Semantic Development 13
“table” followed by a picture of a table) or different items (e.g., the word “table” followed
by a picture of a chair). The logic of this paradigm was that links between pairs of concepts
should interfere with the ability to say that a word for one concept does not denote the
same item as a subsequent picture (e.g., Gellatly & Gregg, 1975). Therefore, co-
occurrence and taxonomic links were measured based on the degree to which they
interfered with participants’ ability to identify a picture as denoting a different item from its
preceding word, relative to unrelated word-picture pairs.
Experiment 3 was designed to provide a more sensitive and graded measure of co-
occurrence and taxonomic links. Specifically, we used a variant of Visual World paradigm
in which we presented pairs of pictures of familiar, unrelated Target items (e.g., bed and
fish), and measured the degree to which participants looked at a given Target over time
upon hearing either a Co-Occur (e.g., pillow or water), Taxonomic (e.g., table or bird), or
Unrelated Prime. Unlike the paradigms used in Experiments 1 and 2, this paradigm does
not provide a single measure of the relatedness between a given pair of concepts, such
as recall accuracy or reaction time. Instead, this paradigm provides a nuanced and
graded measure of semantic relatedness: The degree to which hearing the word for one
concept influences (over time) looking at a picture of the other concept. As attested by
numerous findings, these looking dynamics are sensitive to a variety of relation types,
including extremely weak taxonomic relations (Huettig & Altmann, 2005; Huettig, Quinlan,
McDonald, & Altmann, 2006; Mirman & Graziano, 2012; Mirman & Magnuson, 2009).
Therefore, we measured co-occurrence and taxonomic links based on time course of
looking at each Target when it was accompanied by a Co-Occur versus an Unrelated
Prime, and a Taxonomic versus an Unrelated Prime.
Statistical Regularities Shape Semantic Development 14
Experiment 1
Method
Participants. Informed consent was obtained from parents/guardians of child
participants and from adult participants prior to participation. The sample included 31 4-
5-year-old children (M
age
=4.50 years, SD=1.62 years), and 35 Adults (M
age
=20.16 years,
SD=3.66 years). An additional group of seven children and four adults were tested but
excluded due to either failure to respond on over a third of trials (six children; three adults),
or responding inaccurately on all trials (one child; one adult). An additional eight children
completed practice trials only due to failure to reach the accuracy criterion during these
trials needed to continue to the experiment (see Procedure below). Children were
recruited from families, daycares, and preschools in a metropolitan area of a large
Midwestern US city. Adults were undergraduate students at a large Midwestern public
university in the same city, and they participated in exchange for partial course credit.
These age groups were chosen because: 1) The 4-5 years period is one in which the
nature of sematic knowledge remains the subject of active debate, and 2) Comparing
semantic organization in early childhood to adulthood affords an investigation into
whether early semantic organization is maintained, supplemented, or overwritten by
adulthood.
Selection of Candidate Stimuli. The primary stimuli used in this experiment were
word pairs, with each belonging to one semantic Relatedness condition: Co-Occur,
Taxonomic, or Unrelated. All pairs were selected (as described below) such that pairs in
the Co-Occur condition were words that reliably co-occurred with each other more often
than with other words in child speech input, pairs in the Taxonomic condition were words
Statistical Regularities Shape Semantic Development 15
for concepts from the same category with similar meanings according to a database
composed by lexicographers of words and their definitions (Wordnet, 2010), and pairs in
the Unrelated condition neither reliably co-occurred nor were similar in meaning.
Co-Occurrence Criteria. The first step taken to select pairs in each condition was to
identify a set of words for which lexical norms collected using the MacArthur-Bates
Communicative Development Inventory (MB-CDI) were available from WordBank (an
open database of children's vocabulary development, Frank, Braginsky, Yurovsky, &
Marchman, 2016), and measure their rates of co-occurrence in 25 child speech input
corpora from the CHILDES database (MacWhinney, 2000)
2
. To reduce the computational
expense of measuring co-occurrence rates between these words, some classes of words
(e.g., all sounds such as “moo”) that would a priori not be used as stimuli in this research
were removed from the full set of words, leaving a list of 538 words. Additionally, to ensure
that co-occurrences were measured from child speech input, the CHILDES corpora were
pre-processed to remove all speech produced by the children themselves. Co-
occurrences between these words were then calculated by taking all possible pairs of
words within this set, and calculating how frequently they co-occurred with each other
within a 7-word window
3
across 25 CHILDES corpora. Finally, to account for the fact that
more frequent words co-occur with other words simply by chance, t-scores (Evert, 2008)
were calculated for each word pair using the formula below. This formula captures the
difference between each word pair’s actual, measured co-occurrence frequencies (O),
and the frequency of co-occurrence that would be expected by chance given the
2
The rates of co-occurrence of words in speech input are likely to be similar to the rates with which the
objects to which the words refer co-occur (Sadeghi et al., 2015).
3
This 7-word window was chosen to focus on a span of words that children could plausibly maintain in
memory (e.g., Klem, MelbyLervåg, Hagtvet et al., 2015)
Statistical Regularities Shape Semantic Development 16
frequencies of each word in the pair and the size of the combined corpora (E). The larger
the difference between observed versus expected frequency, the more reliably words in
a pair co-occur:
.  = − 
Candidate word pairs for use in the Co-Occur condition were then selected as pairs
of nouns with t-scores of > 2.5 (following Baayen, Davidson, & Bates, 2008) in which,
according to lexical norms accessed from WordBank, both words were produced by at
least 80% of 36-month-old children (approximately one year younger than children in our
4-year-old sample).
Taxonomic Criteria. Taxonomic relatedness was determined based on both the
membership of concepts in the same taxonomic category (e.g., clothing, foods, vehicles)
and similarity in meaning between their labels. To measure similarity in meaning, we
measured similarity between the definitions of labels for the items from WordNet (a
database of words and their definitions composed by lexicographers). Similarity in
WordNet was chosen as the taxonomic relatedness criterion because it captures the
essence of taxonomic relatedness – i.e., close similarity in meaning – without relying on
judgments of adult participants that may be influenced by non-taxonomic relations
(Wisniewski & Bassok, 1999). In WordNet, nouns are first grouped into sets of synonyms,
which are in turn linked into a hierarchy according to “IS A” and part-whole relations.
Similarity in meaning between pairs of words that label stimuli used in this experiment
was measured using the Resnik similarity measure, i.e., the information content
(specificity) of the word lowest in the WordNet hierarchy within which the pair of words is
subsumed. For example, dog and cat are subsumed within carnivore, whereas dog and
Statistical Regularities Shape Semantic Development 17
kangaroo are subsumed within mammal; because the information content of carnivore is
greater than the information content of mammal (i.e., mammal is more abstract), Resnik
similarity is higher between dog and cat versus dog and kangaroo.
Candidate word pairs for use in the Taxonomic condition were selected as pairs of
nouns with Resnik similarities of >5 in which, as in the selection of candidate word pairs
in other conditions, both words were produced by at least 80% of 36-month-old children
according to WordBank production norms. The rationale of the Resnik similarity criterion
Figure 1. Graphs depicting Resnik similarity between Taxonomic pairs versus other items.
Each graph depicts the Resnik similarity between one item from a Taxonomic pair and: (1)
The other item from the pair (highlighted), (2) Other items from the same taxonomic
category, and (3) Items from other categories. These graphs depict that members of the
same taxonomic category had Resnik similarities greater than five, whereas members of
other categories had similarities substantially lower than five.
Statistical Regularities Shape Semantic Development 18
of >5 is illustrated in Figure 1, which shows that the similarity between Taxonomic pairs
and other items from the same taxonomic category (e.g., clothing) are above 5, whereas
the similarity between items from different taxonomic categories are substantially below
this value.
Unrelated Criteria. Candidate Unrelated word pairs were pairs of nouns that met the
same WordBank production norm criterion as candidates in the Co-Occur and Taxonomic
conditions, with t-scores of < 1.5 and Resnik similarities of < 4.
Composition of Stimulus Set. From the sets of candidate pairs, eight pairs were
selected for each of the three Relatedness conditions (Co-Occur, Taxonomic, and
Unrelated, for a total of 24 pairs) such that: (1) Pairs in the Co-Occur condition did not
meet the Taxonomic Criteria and pairs in the Taxonomic condition did not meet the
Taxonomic Criteria, (2) The mean percentage of 36-month-olds who produced the words
in the pairs according to Wordbank norms was equated across conditions, and (3) No
words appeared in more than one condition (see Table 1 for all 24 pairs, and Appendix A
for t-score and Resnik similarity measures for each pair). Four additional nouns that met
the WordBank production norm criterion were selected to construct pairs used for
demonstration and practice (see Procedure below). All words were recorded by both a
male and a female speaker using an engaging, child-friendly intonation.
Statistical Regularities Shape Semantic Development 19
Design. The relatedness condition was manipulated within subjects, with each pair
presented in Table 1 occurring only in one condition. Because pilot testing indicated that
12 pairs was the maximum number that could be presented to children without producing
floor effects, the total set of 24 pairs (i.e., eight pairs in each of the three Relatedness
conditions) was divided into two Stimulus Sets. Accordingly, each Stimulus Set contained
12 pairs, with four pairs in each of the three Relatedness conditions. In all word pairs,
each word in a pair was randomly assigned to be either the Cue or to-be-remembered
Target. Across conditions, Cue words were presented using the male speaker’s voice,
and Targets using the female’s voice. Additionally, the 12 word pairs in a Stimulus Set
were pseudorandomized into three blocks, such that: (1) Each pair only appeared once
in the entire Stimulus Set, and (2) Each block contained 1-2 pairs from each condition.
The order of these blocks was counterbalanced across participants.
Table 1
Pairs of words used in the Co-Occur, Taxonomic, and Unrelated
conditions in Experiments 1 & 2.
Co-Occur
Unrelated
Bottle Baby
Ball Puzzle
Crayon Frog
Foot Shoe
Pig Bear
Towel Bread
Brush Hair
Horse Bunny
Blocks Cereal
Cup Juice
Carrot Banana
Balloon Tree
Cheese
Mouse
Fork Bowl
Sheep Pancake
Car Street
Popcorn
Fries
Pizza Lion
Soup Spoon
Airplane
Boat
Fish Bed
Milk Cow
Sock Pajamas
Duck Swing
Paper Pencil Chicken Owl Ice
Cream
Bicycle
Note. Only 8 pairs from each condition were used in Experiment 1. A 9
th
pair was added to each condition (bottom row) in Experiment 2.
Statistical Regularities Shape Semantic Development 20
Procedure. In all experiments reported here, participants were presented with
procedures approved by The Ohio State University Institutional Review Board (Protocol
#: 2004B042, Comprehensive protocol for cognitive development research). Adults were
tested in a quiet space in the lab on campus, and children were tested either in a quiet
space in the lab, or at their preschool or daycare. The procedure was similar for adults
and children, with the following exceptions: 1) The instructions were conveyed verbally
by a hypothesis-blind experimenter for children, and as text on a computer screen for
adults, and 2) Children made verbal responses recorded by the experimenter, whereas
adults typed their responses.
To start, participants were introduced to two sock puppets depicted on the computer
screen, named Izzy and Ozzy. Participants were informed that they were going to play a
game with Izzy and Ozzy, in which Izzy and Ozzy would say pairs of words (children were
given an additional explanation about what a “pair” is). The two unrelated Cue-Target
word pairs selected for demonstration/practice were then played sequentially, while
animations depicted one puppet saying the Cue word, and the other saying the Target
word. Next, participants were told that they were going to listen to the word pairs again,
but to pay close attention to the pairs of words that go together, because it would then be
their job to pretend to be Ozzy and either say (children) or type (adults) the word that went
with Izzy’s word. Participants then proceeded to complete practice rounds with the same
two unrelated Cue-Target word pairs. Each practice round consisted of: 1) A “Study”
phase, in which the two word pairs were presented as spoken by Izzy and Ozzy, and 2)
A “Test” phase, in which only the Cue in each pair was presented as spoken by Izzy, and
participants were prompted to either say or type the Target that had been spoken by Ozzy
Statistical Regularities Shape Semantic Development 21
(Figure 2). Participants received feedback about whether their responses were correct or
incorrect. Participants completed up to three practice rounds until they either responded
with the correct Target for both Cues within around, or the experiment was terminated
due to failure to reach this criterion.
Participants then proceeded to complete the three blocks of Cue-Target pairs in the
Stimulus Set to which they had been randomly assigned. Each block followed the same
Study and Test phase format as the practice rounds, with the exception that participants
did not receive feedback about the accuracy of their responses. At the beginning of the
Test phase of each block, participants were encouraged to take their best guess when
they were unsure of the correct answer. The full experiment took approximately 7-10
minutes for adults, and 10-12 minutes for children.
A
B
Figure 2. Schematic of a trial in the Study Phase (A), in which one puppet “says” a
Cue and the other a Target word, and a trial in the Test Phase (B), in which one
puppet says the Cue, and the participant attempts to recall the Target.
Statistical Regularities Shape Semantic Development 22
Results and Discussion
The primary outcome measure of interest for this study was the accuracy with which
participants recalled Target words paired with Cues in each of the three Relatedness
conditions: Co-Occur, Taxonomic, and Unrelated. Responses were scored as accurate
when participants made responses identical to the Target or morphological variants of the
Target (e.g., “spoons” instead of “spoon”). Additionally, three responses (all in children)
in which the correct Target was “street” and the child responded “road” were also scored
as accurate (all reported analyses produce the same results when these responses are
excluded). No other cases of responses synonymous with the Target occurred.
All analyses were conducted in the R environment. Mixed effects models and
corresponding
2
or F-statistics were generated using the lme4 (Bates, Maechler, Bolker,
& Walker, 2015) and car (Fox & Weisberg, 2011) packages, respectively.
Preliminary Analyses: Stimulus Set Comparison. Prior to comparing accuracy in
the Semantic Relatedness conditions, we first tested whether any effect of condition
varied across the two Stimulus Sets in children and adults. Specifically, for data from each
age group, we generated a binomial generalized linear mixed effects model with Accuracy
as the outcome variable, Relatedness condition (Co-Occur, Taxonomic, and Unrelated)
and Stimulus Set (1 versus 2) as fixed effects, and random intercepts for participant and
item. This analysis revealed no significant interaction between Relatedness condition and
Stimulus Set (ps > .09). For all subsequent analyses, we therefore collapsed across
Stimulus Sets.
Primary Analyses. Memory accuracies by age and condition are presented in Figure
3. To test the relative influences of Semantic Relatedness conditions on recall accuracy,
Statistical Regularities Shape Semantic Development 23
we generated an omnibus binomial generalized linear mixed effects model with Accuracy
as the outcome variable, Relatedness condition (Co-Occur, Taxonomic, and Unrelated)
and Age group (children and adults) as fixed effects, and random intercepts for participant
and item. This analysis revealed main effects of both Relatedness condition (
2
(2)=20.35,
p<.001) and Age group (
2
(1)=15.74, p<.001) that were qualified by an interaction
between them (
2
(2)=9.38, p=.009).
To investigate the interaction between Relatedness condition and Age group, we
compared the effects of the different Relatedness conditions in each Age group.
Relatedness conditions in Each Age Group. To compare the effects of the
Relatedness conditions in each Age Group, for each Age group, we first generated a
binomial generalized linear mixed effects model with Accuracy as the outcome variable,
Relatedness condition as a fixed effect, and random intercepts for participant and item.
These models revealed significant effects of Relatedness condition in each age group (ps
< .01) (Figure 3). To conduct pairwise comparisons of the Relatedness conditions in each
age group, we re-generated the model for each age with each of the Relatedness
conditions as the reference level, and applied Bonferroni-adjustments to the resulting p-
values.
In children, these analyses revealed significant differences between the Co-Occur
and both Unrelated and Taxonomic conditions (ps < .001), but no difference between the
Taxonomic and Unrelated conditions (p > .99). In adults, these analyses revealed a
significant difference between the Co-Occur and Unrelated conditions (p = .003), the
Taxonomic and Unrelated conditions (p=.04), and no significant difference between Co-
Occur and Taxonomic conditions (p>.99).
Statistical Regularities Shape Semantic Development 24
Individual Differences. The results of the primary analyses suggest that Co-
Occurrence links manifested in semantic organization in both children and adults,
whereas Taxonomic relatedness manifested only in adults’ semantic organization.
However, it is important to highlight that the lack of a taxonomic influence in young
children observed in the present experiments is a null finding, from which strong
conclusions cannot be drawn. For example, an influence of taxonomic relatedness may
have been present, but was too small in magnitude and/or transpired in too few children
to detect. We therefore investigated this possibility using a qualitative analysis of the
magnitudes of Co-Occur and Taxonomic relatedness effects within individuals in both the
child and adult samples.
A
B
Figure 3. Proportion accurate in children (panel A) and adults (panel B) in the three
Relatedness conditions. Error bars represent standard errors of the means.
Statistical Regularities Shape Semantic Development 25
In these analyses, we quantified the magnitude of both Co-occurrence and
Taxonomic effects for each participant by calculating both a Co-Occur and a Taxonomic
Difference Score based on the difference between each of these conditions and the
Unrelated condition (such that a Difference Score of 0 for a given condition indicates no
influence of the condition on behavior). The densities of the distributions of Difference
Scores in each age group in each experiment are depicted in Figure 4. As in the primary
analyses, these distributions show an influence of Co-occurrence relatedness in children,
and both Co-occurrence and Taxonomic relatedness in adults. However, these
distributions also suggest an influence of Taxonomic relatedness was present in children,
but both occurred in fewer participants and tended to be smaller in magnitude than the
influence of Co-Occurrence relatedness.
Figure 4. Kernel densities for Co-Occur and Taxonomic Difference Scores in Experiment
1. Difference Scores were calculated by comparing the Co-Occur and Taxonomic
conditions to the Unrelated condition, such that larger values correspond to larger
influences of a given condition (i.e., greater improvement in accuracy).
Statistical Regularities Shape Semantic Development 26
Summary. The results of Experiment 1 revealed a substantial influence of co-
occurrence regularities in both young children and adults, such that co-occurrence
between to-be-remembered Cue and Target word pairs facilitated subsequent recall. In
contrast, taxonomic relatedness did not significantly affect recall in children. Instead, the
influence of taxonomic relatedness transpired only in adults. However, our quantification
of co-occurrence and taxonomic effects within individuals adds nuance to this pattern.
Specifically, this qualitative analysis both corroborates these results, and suggests that
taxonomic contributions, rather than being totally absent in children, were instead present
but too weak and uncommon to produce significant effects at the group level.
These results highlight the role of co-occurrence in semantic organization throughout
development. Moreover, these results also suggest that over development, new (and
perhaps more advanced) taxonomic organization increasingly supplements co-
occurrence rather than replaces it (see also Supplementary Materials for quantifications
of positive correlations between co-occurrence and taxonomic effects, consistent with the
proposal that taxonomic relations build upon co-occurrence). To examine the
generalizability of this finding, we conducted Experiment 2.
In Experiment 2, we used an entirely different paradigm to investigate participants’
sensitivity to co-occurrence and taxonomic relatedness. In this paradigm, participants
were presented with word-picture pairs in which the picture either did or did not depict the
item referred to by the word (e.g., the word “lion” followed by a picture of a lion or the
word “bottle” followed by a picture of a baby). In contrast to examining whether
relatedness improved performance (as was done in Experiment 1), we used this paradigm
to measure the degree to which co-occurrence or taxonomic relatedness between word-
Statistical Regularities Shape Semantic Development 27
picture pairs interfered with participants’ ability to indicate that the picture did not depict
the item labeled by the word.
Experiment 2
Method
Participants. Informed consent was obtained from parents/guardians of child
participants and from adult participants prior to participation. The sample included 41 4-
5-year-olds (M
age
=4.05 years, SD=1.71 years) and 42 Adults. Two additional children
were tested but excluded due to mean reaction times more than two standard deviations
above the mean reaction time for this age group. An additional three children completed
practice trials only due to failure to reach the accuracy criterion needed to continue to the
experiment (see Procedure below). Children were recruited from families, daycares, and
preschools in a metropolitan area in a Midwestern US city. Adults were recruited from the
undergraduate population at a public university in the same city and participated in
exchange for partial course credit.
Stimuli and Design. The primary stimuli were similar to those used in Experiment 1,
with the following changes. First, we added several pairs to those used in Experiment 1.
Specifically, we added one pair to the Co-Occur, Taxonomic, and Unrelated conditions,
for a total of 9 pairs in each condition. Additionally, from the list of nouns not used in the
Co-Occur, Taxonomic, or Unrelated conditions, an additional 24 nouns that met the
production by 80% of 36-month-olds WordBank criterion were selected for use in an
Identical condition in which a “pair” consisted of a word and a picture depicting the same
thing (e.g., the word “lion” followed by a picture of a lion). The Co-Occur, Taxonomic,
Statistical Regularities Shape Semantic Development 28
Unrelated, and Identical pairs each appeared once in the experiment, for a total of 51
pairs. These pairs were pseudo-randomized prior to the experiment such that no more
than two pairs from the same condition appeared consecutively. An additional 18 nouns
that met the WordBank production norm criterion were also selected to appear as
demonstration and practice stimuli (see Procedure below).
Second, whereas the stimuli were divided into separate Stimulus Sets in Experiment
1, all stimuli were presented to all participants in this experiment. Finally, to eliminate
potential effects of perceptual similarity between members of the same pair, both spoken
words and pictures were generated for all words, such that a spoken word was used for
one member of the pair and a picture for the other member. Specifically, for each pair,
one word was randomly assigned to appear in the experiment as a spoken word, and the
other was assigned to appear as a picture. The spoken word version was recorded by a
male speaker using an engaging, child-friendly intonation, and the picture version was a
color photograph of the item isolated on a white background (resized to 276x276pi).
As in Experiment 1, the relatedness condition varied within subjects. In addition, also
as in Experiment 1, each word pair was presented only once.
Procedure. Adults were tested in a quiet space in the lab on campus, and children
were tested either in a quiet space in the lab, or at their preschool or daycare. The
procedure was similar for adults and children, with the exceptions that: 1) The instructions
were conveyed verbally by a hypothesis-blind experimenter for children, and as text on a
computer screen for adults, and 2) Children chose response options using a touchscreen,
whereas adults used a mouse.
Statistical Regularities Shape Semantic Development 29
To start, participants were instructed that they were going to play a game in which
they would hear a word and then see a picture, and that their job was to identify whether
the picture was “of the same thing” as the word. Participants were then instructed to click
(adults) or touch (children) a smiley face depicted on the bottom of the screen if the picture
was of the same thing as the word. Two Demonstration trials were then presented. In
these trials, a word was followed by a picture of the same thing (e.g., the word “pretzel”
followed by a picture of a pretzel), and the smiley face was highlighted as the correct
response. Subsequently, participants were instructed to click or touch a frowny face also
depicted at the bottom of the screen if the picture was not of the same thing as the word,
and then shown Demonstration trials using two unrelated word-picture pairs (e.g., the
word “zebra” followed by a picture of scissors). Participants then proceeded to complete
eight Practice trials composed of an equal number of word-picture pairs in which the
picture was of the same thing as the word, and pairs in which the picture was of an item
unrelated to the word. Participants were encouraged to respond as quickly and accurately
as possible. The smiley and frowny face response options only appeared 250ms after the
onset of the picture, and remained on the screen for 6 seconds, to impose a time limit on
the window within which responses could be made. Participants received corrective
feedback after each trial telling them whether they were correct, incorrect, or too slow (if
they failed to respond during the time limit). If a participant failed to reach a criterion of 5
out of 8 trials correct, the Practice trials repeated up to two additional times until the
criterion was reached. If a participant failed to reach the criterion after three rounds of
Practice trials, the experiment was terminated for that participant.
Statistical Regularities Shape Semantic Development 30
Participants then proceeded to complete the experimental trials, in which items from
the Co-Occur, Taxonomic, Unrelated, and Identical conditions were presented in a
pseudo-randomized order, such that no more than two pairs from the same condition
were presented consecutively. These trials followed the same format as Practice trials,
with the exceptions that the response options remained on the screen until participants
made a response (i.e., no upper time limit was imposed), and no feedback was provided.
The full experiment took approximately 4-5 minutes for adults, and 5-8 minutes for
children.
Results and Discussion
In the experiment, accurate responses were those in which participants responded
that the word and picture were of the same thing (henceforth referred to as “yes”
responses) in the Identical condition, or that the word and picture were not of the same
thing (henceforth referred to as “no” responses) in all other conditions. Prior to conducting
hypothesis-testing analyses, we first determined that both children and adults understood
the task: In both age groups, overall response accuracies for both “yes” and “no”
responses were significantly above chance (Children: M
yes
= 83.94%, M
no
= 88.41%;
Adults: M
yes
=97.62%, M
no
= 98.54%; all ps < .001).
Primary Analyses. Reaction times by age and condition are presented in Figure 4.
Our primary measure of interest was how much more difficult it was for participants to
make accurate “no” responses to non-identical word-picture pairs in the Co-Occur and
Taxonomic conditions compared to the Unrelated condition. We measured comparative
difficulty using reaction time (log-transformed for analyses) in the three conditions. Such
Statistical Regularities Shape Semantic Development 31
interference effects in the Co-Occur or Taxonomic conditions were taken as evidence that
a participant was sensitive to the respective relation.
To test the relative influences of the Relatedness conditions on reaction time, we
generated an omnibus linear mixed effects model with Reaction Time as the outcome
variable, Relatedness condition (Co-Occur, Taxonomic, and Unrelated) and Age group
(children and adults) as fixed effects, and a random intercept for participant. This analysis
revealed main effects of both Relatedness condition, F(2,1977.26)=21.89, p< .001 and
Age, F(1, 81.01)=248.66, p<.001, which were qualified by an interaction, F(2,1977.51)=
11.29, p<.001.
Relatedness conditions in Each Age Group. To investigate the interaction between
Relatedness condition and Age group, for each Age group, we first generated a linear
A
B
Figure 5. Reaction Times in children (panel A) and adults (panel B) in the three
Relatedness conditions. Error bars represent standard errors of the means. The y-axes
for the two age groups are different because children’s reaction times were substantially
longer than those of adults.
Statistical Regularities Shape Semantic Development 32
mixed effects model with Reaction Time as the outcome variable, Relatedness condition
as a fixed effect, and participants as a random effect. These models revealed significant
effects of Relatedness condition in each age group (ps < .05).
To conduct pairwise comparisons of the Relatedness conditions in each age group,
as in analyses for Experiment 1, we re-generated the model for each age group with each
of the Relatedness conditions as the reference level, and applied Bonferroni-adjustments
to the resulting p-values. In children, these analyses revealed significant differences
between the Co-Occur and both Unrelated and Taxonomic conditions (ps < .001), but no
difference between the Taxonomic and Unrelated conditions (p > .99). In adults, these
analyses revealed a significant difference between the Co-Occur and Unrelated
conditions (p = .014), no significant difference between Co-Occur and Taxonomic
conditions (p=.352), and no significant difference between the Taxonomic and Unrelated
conditions (p=.630). However, there was a numerical trend for longer reaction times in
both Co-Occur versus the Taxonomic condition, and for the Taxonomic versus the
Unrelated Condition (M
co-occur
= 974ms, M
taxonomic
= 945ms, M
unrelated
= 928ms).
Individual Differences. We supplemented our primary analyses by following the
same approach as in Experiment 1 to quantifying the magnitudes of Co-Occur and
Taxonomic relatedness effects within individuals in both the child and adult samples. As
shown in Figure 6, as in Experiment 1, this analysis both corroborates the results of our
primary analyses, and suggests that taxonomic relatedness effects were present but too
weak and uncommon in children to reach significance at the group level.
Statistical Regularities Shape Semantic Development 33
Summary. The results of this experiment provided further evidence for a substantial
sensitivity to co-occurrence regularities that manifested in young children, and continued
into adulthood. Specifically, in both age groups, participants found it more difficult to
identify when a picture did not depict the same thing as a preceding word if the word and
the picture’s label reliably co-occur in linguistic input.
In contrast, sensitivity to taxonomic relatedness in this task did not reach significance
in young children. As in Experiment 1, our qualitative analysis of individual co-occurrence
and taxonomic effects suggests that the absence of the taxonomic effects at the group
level in children was due to the weakness and rarity of these effects in children, rather
than to their complete absence. In adults, although the influence of taxonomic relatedness
Figure 6. Kernel densities for Co-Occur and Taxonomic Difference Scores in Experiment
2. Difference Scores were calculated by comparing the Co-Occur and Taxonomic
conditions to the Unrelated condition, such that larger values correspond to larger
influences of a given condition (i.e., greater slowing of reaction time).
Statistical Regularities Shape Semantic Development 34
was not significantly smaller than the influence of co-occurrence, responses to taxonomic
pairs also did not significantly differ from responses to unrelated pairs. This replication of
the contribution of co-occurrence in children and adults using two very different paradigms
underscores the significance of sensitivity to co-occurrence regularities in relational
knowledge across development (see also Supplementary Materials for quantifications of
positive correlations between co-occurrence and taxonomic effects, consistent with the
proposal that taxonomic relations build upon co-occurrence).
The purpose of Experiment 3 was to both test the generalizability of these patterns to
another very different paradigm, and to gain a more sensitive and nuanced measure of
co-occurrence and taxonomic links. Specifically, our qualitative analyses for both
Experiments 1 and 2 suggested taxonomic links in young children that were too weak to
reach significance at the group level. Therefore, in Experiment 3, we used a paradigm
that has been shown to yield a sensitive and graded measure of even weak semantic
relations: The Visual World paradigm.
In the Visual World paradigm, participants view items (typically pictures) while hearing
linguistic input, such as a word. Numerous studies have provided evidence that
individuals tend to look at pictures that are semantically related to the words that they
hear (e.g., Huettig & Altmann, 2005; Mirman & Magnuson, 2009). Accordingly, the degree
to which hearing a word for one concept (e.g. cat) prompts looking at a picture of a
semantically related concept (e.g., dog) can serve as a measure of the degree to which
the concepts are linked in an individual’s semantic knowledge. A similar tendency has
been observed in infants in preferential looking paradigms (Arias-Trejo & Plunkett, 2009;
Statistical Regularities Shape Semantic Development 35
Bergelson & Aslin, 2017) suggesting that measures of this looking behavior are
appropriate for a wide developmental age range.
Critically, unlike the paradigms used in Experiments 1 and 2, this paradigm does not
yield only a single snapshot measurement of relatedness between two concepts. Instead,
it measures the degree to which one concept (presented as a word), activates another
concept (presented as a picture). Moreover, it measures how this degree of activation
unfolds over time in the milliseconds following the presentation of the word. This paradigm
therefore yields a graded, nuanced measure that has been shown to be sensitive to even
weak taxonomic relations (Mirman & Magnuson, 2009).
Experiment 3
Overview
For Experiment 3, we developed a variant of the Visual World paradigm with key
characteristics designed to probe the degree to which words activate co-occurring and
taxonomically related concepts directly. In this paradigm, participants heard “Prime”
words while freely visually inspecting visual displays containing two “Target” pictures of
unrelated familiar items (e.g., bed and fish). A given pair of Target pictures always
appeared with each other, and never with other items. Across presentations of a given
Target pair, we varied whether the Prime word was: (1) A Co-Occur Prime that co-
occurred with one of the Targets (e.g., pillow or water), (2) A Taxonomic Prime that was
taxonomically related with one of the Targets (e.g., table or bird), or (3) An Unrelated
Prime that was unrelated to both Targets. Following presentation of the Prime,
participants freely viewed the Targets for 2000ms. We measured the activation of a
Statistical Regularities Shape Semantic Development 36
Target by a Prime based on the degree to which participants looked more at a given
Target over time when accompanied by a Co-Occur or Taxonomic versus an Unrelated
Prime.
It is worth highlighting two characteristics that distinguish this version of the Visual
World paradigm from the ways in which this paradigm has typically been implemented in
prior research. First, whereas prior approaches have manipulated semantic relatedness
by manipulating the pictures that appear with a given Prime word (e.g., presenting “bed”
with either a picture of a pillow or a chair), our version manipulated the Prime word that
was presented with a given pair of pictures. This approach allowed us to measure the
temporal dynamics with which the concepts depicted by the Target pictures were
activated upon hearing different Primes while keeping the pictures themselves constant,
and therefore avoiding contamination from visual salience, visual interest, and so on.
Second, in trials in which participants heard a Co-Occur, Taxonomic, or Unrelated Prime,
participants did not complete a task, and instead freely viewed the Targets. This
characteristic kept our measure of semantic relatedness between Primes and Targets
implicit, as in Experiment 1 and 2. Instead, these trials were interspersed with trials of a
cover task, in which participants heard the word “yellow” or “blue”, and clicked a button
on a button box of the corresponding color.
Method
Participants. Informed consent was obtained from parents/guardians of child
participants and from adult participants prior to participation. The sample included 36 4-
5-year-olds (M
age
=4.43 years, SD=0.32 years) and 37 Adults. Children were recruited
from families, daycares, and preschools in a metropolitan area in a Midwestern US city.
Statistical Regularities Shape Semantic Development 37
Adults were recruited from the undergraduate population at a public university in the same
city and participated in exchange for partial course credit.
Stimuli. The primary stimuli were similar to those used in Experiments 1 and 2. The
primary difference in this experiment was that, instead of separate sets of Co-Occur,
Taxonomic, and Unrelated pairs, we constructed sets in which a Target was combined
with both a Co-Occur Prime and a Taxonomic Prime. Further, we organized these sets
Table 2
Pair Sets in Experiment 3. Targets were presented as pictures, and Primes as
words.
Pair Set
Target
Co-Occur Prime
Tax Prime
Unrelated Prime
1
Car
Street
Bicycle
Stick
Nose Tissue Tongue
2
Airplane
Sky
Boat
Bathtub
Cheese Mouse Ice Cream
3
Horse
Cowboy
Frog
Cloud
Pizza Oven Chocolate
4
Bread
Knife
Muffin
Button
Foot Shoe Head
5
Fish
Water
Bird
Train
Bed Pillow Table
6
Cow
Milk
Tiger
Book
Leg Pants Finger
7
Cereal
Breakfast
Pancake
Shovel
Monkey Zoo Squirrel
8
Soup
Spoon
Juice
Telephone
Coat Zipper Sweater
9
Bottle
Baby
Bowl
Door
Apple Tree Grapes
10
Corn
Chicken
Pumpkin
Star
Sock Foot Hat
11
Owl
Moon
Turtle
Nail
Bucket
Hose
Jar
Statistical Regularities Shape Semantic Development 38
into pairs (“Pair Sets”) in which: (1) The Targets in the Pair Set were both unrelated and
approximately equivalently familiar (i.e., were produced by a similar percentage of 36-
month-old children according to production norms), and (2) The Primes for one Target in
a Pair Set were unrelated to the other Target (Table 2 and Appendix B). Each Pair Set
additionally included one Unrelated Prime that was unrelated to both Targets.
Primes were presented as words recorded in the same manner as in Experiments 1
and 2. We additionally recorded the words “yellow” and “blue” for use in cover task filler
trials (see Procedure). Targets were presented as pictures each subtending
approximately 5.3 of visual angle.
Finally, the total number of items was expanded in this experiment following the same
co-occurrence and taxonomic criteria as in Experiments 1 and 2, for a total of 22 sets
organized into 11 Pair Sets
4
Design. As in Experiments 1 and 2, the Relatedness condition (Co-Occur,
Taxonomic, and Unrelated) varied within subjects, and Age varied between subjects.
Within a block of trials, there were a total of 88 trials comprised of 22 trials of each of the
following four types: (1) Co-Occur (each of the 11 Target pairs were presented with the
two Co-Occur Primes from their Pair Set), (2) Taxonomic (each of the 11 Target pairs
were presented with the two Taxonomic Primes from their Pair Set), (3) Unrelated (each
of the 11 Target pairs were presented twice with the Unrelated Prime from their Pair Set),
and (4) Filler (each of the 11 Target pairs were presented with the words “yellow” and
“blue”). Children completed a single block of trials, and adults completed two blocks.
4
To generate this larger stimulus set, we relaxed the familiarity criterion such that words only needed to be
produced by at least 55% of 36-month-olds. However: (1) The majority (86%) of words still met the 80%
criterion used in Experiments 1 and 2, and (2) The average production norm value was equated across Co-
Occur, Taxonomic, and Unrelated Primes (all produced by ~89% of 36-month-olds).
Statistical Regularities Shape Semantic Development 39
Apparatus. This experiment used an EyeLink Portable Duo eye tracking system that
measures eye gaze by computing the pupil-corneal reflection at a sampling rate of 500Hz.
We additionally constructed a non-functional “button box” with yellow and blue buttons for
use in the cover task that participants completed during the experiment (see Procedure).
Procedure. Adults were tested in a quiet space in the lab on campus, and children
were tested either in a quiet space in the lab, 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).
Following calibration of the eye tracker, participants began a practice session of a
cover task. The purpose of this cover task was to keep participants engaged in looking at
the screen and listening to the words, but consisted only of filler trials that were not
Figure 7. Timing of events in Experiment 3 trials. Note: The trial ended 2000ms post-
Prime Onset in Experimental trials only. In cover task filler trials, it ended either:
When terminated by the experimenter upon observing the participant clicking the
yellow or blue button, or after 5000ms.
Statistical Regularities Shape Semantic Development 40
analyzed. Specifically, to perform the cover task, participants were given the non-
functional button box and told that they would see two pictures: One on a yellow
background on the left, and one on a blue background on the right. They would then hear
either the word “yellow”, or the word “blue”, such that their job was to click the yellow
button on the button box if they heard “yellow”, and the blue button if they heard “blue”.
The practice session consisted of 11 trials of this cover task. In each trial, the two pictures
on the screen were the two Targets from one of the 11 Pair Sets. In cover task filler trials
only, the experimenter terminated the trial upon observing the participant clicking one of
the buttons. The timing of events in cover task and subsequent experimental trials is
depicted in Figure 7.
Following completion of the practice session, participants were informed that they
would continue to play the same game, but that it would get “a bit tricky”, because
sometimes they would hear a word that was neither blue nor yellow. Participants were
instructed to not click either of the buttons if this occurred.
Participants then proceeded to complete either one block of trials (children) or two
blocks (adults). On each trial, the two pictures on the screen were the two Targets from
one of the Pair Sets. To create the three Relatedness conditions, in experimental trials,
the word was either: (1) The Co-occurrence Prime for one of the Targets, (2) the
Taxonomic Prime for one of the Targets, or (3) The Unrelated Prime for the Targets.
These experimental trials were randomly ordered and interspersed with the above
described cover task filler trials, in which the word was either “yellow” or “blue”. The pairs
of Targets in the Pair Set were each presented twice in filler trials, once with the word
“yellow” and once with “blue”. In combination with the experimental trials, the Targets from
Statistical Regularities Shape Semantic Development 41
each Pair Set were therefore presented a total of 8 times within a block (twice in each of
the three Relatedness conditions and twice in filler trials), within which the locations of the
Targets were counterbalanced. The full experiment took approximately 10-12 minutes for
children and adults.
Using this procedure, we measured the degree to which the looking dynamics for the
two Target pictures in a Pair Set varied according to the relation between each Target
and the Primes. The fact that each pair of Target pictures was always presented together
across the Prime relatedness conditions allowed us to control for effects of visual
features, salience, similarity, etc. while measuring these looking dynamics.
Results and Discussion
To test the contributions of co-occurrence and taxonomic relatedness in children and
adults, the data from this experiment were used to compare the time course of looking at
Targets accompanied by Co-Occurring or Taxonomic Primes versus Unrelated Primes in
children and adults. To conduct this comparison, we first processed the raw eye tracking
data to generate 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 within neither AOI. We first removed data
from the 500ms prior to the onset of the word, 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” value for each Target
in each time bin in the Co-Occurring, Taxonomic, and Unrelated Prime conditions. This
Statistical Regularities Shape Semantic Development 42
Target Dwell Time value captured the amount of time spent looking at the Target in each
time bin when it was accompanied by a Co-Occurring, Taxonomic, or an Unrelated Prime.
These values were used to test whether the time course of looking at a Target differed
when accompanied by a Co-Occur or Taxonomic versus an Unrelated Prime in children
and adults (for analyses of the proportion of dwell time for each Target, out of the total
dwell time to both Targets in a Pair Set, see Supplemental Materials).
Difference from Unrelated. To test for differences in the degree to which the Co-
Occur versus Taxonomic Prime conditions deviated from the Unrelated Prime condition,
we calculated a “Difference from Unrelated” value for each Target in each time bin. 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 Co-Occur condition, and the Target
Dwell Time in the Taxonomic condition. The Difference from Unrelated value therefore
captures the degree to which participants looked more at each Target in each time bin
when it was accompanied by a Co-Occur or a Taxonomic Prime than when it was
accompanied by an Unrelated Prime (for comparable analyses of proportion of Target
looking in each time bin, see Supplemental Materials).
Analysis Approach. We followed the Growth Curve Analysis (GCA) approach
developed by Mirman and colleagues (Mirman, Dixon, & Magnuson, 2008) to analyze our
data. The GCA approach involves the generation of hierarchical mixed effects models,
starting with a “base” model that captures how looking behavior changes over time
overall, without considering variation across individuals or experimental manipulations.
First, the intercept captures the average value of the outcome variable. In addition, the
base model also includes a linear term that captures monotonic changes in the value of
Statistical Regularities Shape Semantic Development 43
the outcome variable over time, and a quadratic term that captures the sharpness of the
peak in looking over time. Finally, cubic and quartic terms capture changes in the
asymptotic tails of the outcome variable change over time that are not typically informative
about the influences of experimental manipulations (Mirman et al., 2008).
To analyze the effects of experimental manipulations, the base model is
supplemented with: Fixed effects of experimental manipulations, random intercepts for
participants (and items if appropriate), and random slopes for effects of experimental
manipulations within participants (and items, if appropriate). The interpretation of
significant fixed effects on the model terms are as follows: (1) Effects on the intercept
capture overall effects collapsed across the entire time period on the outcome variable;
(2) Effects on the linear term capture effects on the rate of linear change in the outcome
variable, similar to linear regression; and (3) Effects on the quadratic term capture effects
on the sharpness of the peak with which the outcome variable increases and then
decreases (or decreases then increases).
Target Dwell Time Analysis. We first tested whether the temporal dynamics of
looking at Targets differed when accompanied by Co-Occur 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 children and adults that
both supplemented the base model with a fixed effect of Relatedness condition (with
Unrelated as the reference level to which Co-Occur and Taxonomic were compared).
These models additionally included random intercepts for participant and item, and
random slopes for the effect of Relatedness condition within participants and within items.
Statistical Regularities Shape Semantic Development 44
The parameter estimates and their significance levels are reported in Table 3. Both
children and adults looked more overall at a given Target when they heard either a Co-
Occur or a Taxonomic versus an Unrelated Prime (as shown by significant effects on the
Intercept). Co-Occur 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 (see Supplemental Materials for similar
results from analyses of Target dwell proportions).
However, this analysis does not reveal the relative contributions of Co-Occur versus
Taxonomic Primes. To compare the contributions of co-occurrence and taxonomic
relatedness, in the following analysis, we compared Difference from Unrelated in the Co-
Occur and Taxonomic conditions.
Table 3
Results of growth curve analysis of Target Dwell Times. Parameter estimates are for the Co-
Occur and Taxonomic conditions relative to the Unrelated condition. Non-significant
parameter estimates are in italics.
Co-Occur Taxonomic
Model Term
Age Group
Est. (SE)
p
Est. (SE)
p
Intercept Child 9.599 (1.852) <.001 6.389 (1.852) <.001
Linear Child 31.351 (5.603) <.001 13.229 (5.603) .020
Quadratic Child -4.702 (5.057) .355 -9.130 (5.057) .075
Intercept Adult 8.986 (2.605) <.001 6.768 (2.605) .011
Linear Adult 21.462 (7.149) .003 14.384 (7.149) .047
Quadratic Adult -23.151 (5.097) <.001 -18.836 (5.097) <.001
Statistical Regularities Shape Semantic Development 45
Difference from Unrelated Analysis. This analysis tested whether there was a
difference in the degrees to which the Co-Occur versus Taxonomic conditions deviated
from the Unrelated condition in children and adults. Specifically, we generated separate
models of Difference from Unrelated values for children and adults that both
supplemented the base model with a fixed effect of Relatedness condition (Co-Occur and
Taxonomic only), random intercepts for participant and item, and random slopes for the
effect of Relatedness condition within participants and within items. Figure 8 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 4. In
children, Co-Occur Primes produced grater rates of increased looking at Targets (relative
to Unrelated Primes) than Taxonomic Primes. In contrast, in adults, no such differences
Figure 8. Difference from Unrelated values in the Co-
Occur (red) and Taxonomic
(blue) conditions in Children and Adults, plotted with lines depicting the fitted values
from the models. Error bars depict standard errors of the mean.
Statistical Regularities Shape Semantic Development 46
were observed: Co-Occur and Taxonomic Primes affected looking at Targets relative to
Unrelated Primes to equivalent extents (for similar results from analyses of Target
proportions, see Supplemental Materials).
Summary. The results of this experiment provided further nuance to our picture of
the developmental trajectory of semantic organization. First, this experiment revealed that
an influence of co-occurrence that persisted from early childhood to adulthood,
corroborating results from Experiments 1 and 2.
Critically, this experiment revealed an influence of taxonomic relatedness that was
initially weaker than co-occurrence in young children, but became similar in magnitude to
co-occurrence by adulthood. This result explicitly captures and quantifies the
developmental trajectory suggested by our analyses of individual differences in
Experiments 1 and 2, in which taxonomic relatedness in the course of development
supplements co-occurrence-based links.
Table 4
Results of growth curve analysis of Difference from Unrelated.
Parameter estimates are for the Co-Occur relative to the
Taxonomic condition. Non-significant parameters are in italics.
Co-Occur versus Taxonomic
Model Term
Age Group
Est. (SE)
p
Intercept Child 3.210 (1.970) .113
Linear Child 18.122 (5.938) .004
Quadratic Child 4.428 (5.094) .390
Intercept Adult 2.217 (2.493) .378
Linear Adult 7.078 (7.541) .352
Quadratic Adult -4.315 (5.169) .409
Statistical Regularities Shape Semantic Development 47
General Discussion
Across experiments that used three different paradigms to yield implicit measures of
semantic organization, we observed substantial effects of co-occurrence in both young
children and adults. In contrast, the data suggest that taxonomic relatedness increasingly
supplements co-occurrence with development. Importantly, due to the implicit nature of
the measures of semantic organization used in these experiments, this developmental
pattern is unlikely to be attributable to other developmental changes, such as
improvement in explicit reasoning abilities.
These findings arbitrate between the predictions of different accounts of semantic
organization development. First, the evidence for a continued contribution of co-
occurrence to semantic organization in adults is inconsistent with Restructuring accounts,
which predict that early-emerging organization based on environmental input (such as co-
occurrence) is later overwritten by taxonomic relations. Second, the substantial
contributions of co-occurrence to semantic organization throughout development suggest
that accounts that do not posit any role for co-occurrence, including both Taxonomic Bias
and Featural Learning accounts, are at best incomplete. Specifically, although the
sources of input to semantic organization highlighted by these accounts – e.g., labels in
Taxonomic Bias accounts and features in Featural Learning accounts may indeed
contribute to semantic organization, co-occurrence regularities also appear to play a key
role that these accounts overlook.
The present findings are instead most consistent with a recent mechanistic account
proposed by Sloutsky et al. (2017). According to this account, co-occurrence contributes
to semantic organization from early in development onward because it is directly
Statistical Regularities Shape Semantic Development 48
observable from environmental input. Taxonomic relations then increasingly come to
contribute to semantic organization as they are derived from regularities with which
different labels share patterns of co-occurrence with each other (e.g., members of the
same taxonomic category such as spaghetti and pie share each other’s patterns of co-
occurrence with fork, plate, etc.). The developmental trajectory predicted by this account,
in which co-occurrence contributes to semantic organization throughout development and
is gradually supplemented by taxonomic relations, was corroborated by the results of the
present experiment.
In principle, other, as-of-yet unproposed accounts could also explain the present
findings as the result of two entirely separate processes for forming co-occurrence-based
and taxonomic relations that develop asynchronously. For example, the more gradual
emergence of taxonomic relations might be interpreted as resulting from a gradually-
emerging sensitivity to the features that members of taxonomic categories share (e.g.,
Sloutsky, 2010; Smith & Heise, 1992). Alternately, the gradual emergence of taxonomic
relations might be driven by learning both words such as “animal”, “clothes”, “furniture”
etc., and to infer that these denote stable taxonomic categories (e.g., Fulkerson &
Waxman, 2007; Gelman & Coley, 1990; Gelman & Markman, 1986). However, regardless
of the theoretical framework within which they are interpreted, the findings nonetheless
underline the importance of incorporating a key role for co-occurrence regularities in any
account of semantic development.
However, the trajectory of semantic organization development cannot be inferred
from the present experiments alone. To contextualize these findings, we next evaluate
the degree to which this developmental trajectory is consistent with evidence from prior
Statistical Regularities Shape Semantic Development 49
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. Finally, we discuss potential mechanistic explanations
for the developmental trajectory observed in the present experiments that represent
targets for future research.
Developmental Trajectories Observed in Present and Prior Research
Contribution of Co-Occurrence. Across the three experiments, we observed
significant contributions of co-occurrence to semantic organization from early childhood
into adulthood. In both young children and adults, co-occurrence: (1) Improved recall of
word pairs, (2) Interfered with the ability to identify a picture as not of the same thing as
a preceding word, and (3) Guided the dynamics of looking behavior. In this section, we
evaluate these findings in the context of prior research. Although a role for co-occurrence
throughout semantic organization development has been overlooked or posited to be
transient in the majority existing accounts, the evidence supporting this role provided in
the present experiments is consistent with many prior findings.
First, our evidence that co-occurrence contributes to semantic organization
throughout development is consistent with numerous findings from statistical learning
research. Specifically, multiple statistical learning studies have provided evidence that a
sensitivity to co-occurrences between inputs in many domains, including speech sounds,
acoustic non-speech sounds, and visual objects (e.g., Bulf et al., 2011; Samuelson &
Smith, 1999), emerges in infancy and persists into adulthood. Moreover, beyond being
consistent with this prior evidence, the present findings build upon it by suggesting that
Statistical Regularities Shape Semantic Development 50
sensitivity to co-occurrence regularities also contributes to the domain of semantic
organization.
Second, the present findings corroborate evidence from numerous studies with
children (e.g., Blaye et al., 2006; Lucariello et al., 1992; Walsh et al., 1993) and a handful
of studies with adults (Lin & Murphy, 2001; Murphy, 2001) for the presence of links in
semantic organization that may be derived 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 semantic organization.
Contribution of Taxonomic Relations. Taken together, the results of the three
present experiments suggested that an influence of taxonomic relatedness came to
supplement co-occurrence with development. Specifically, Experiments 1 and 2 did not
detect significant effects of taxonomic relations at the group level in young children, and
instead only detected weak and uncommon effects within individual children. Experiment
3 did detect an influence of taxonomic relations within young children as a group due to
its use of a more sensitive, graded measure, but as in Experiments 1 and 2, this influence
was weaker than the influence of co-occurrence. Across experiments, similar effects of
taxonomic relations and co-occurrence were only observed in adults. Here, we evaluate
this developmental trajectory in the context of prior research.
The degree to which taxonomic relations contribute to semantic organization at
various points in development has been the subject of extensive prior research that has
yielded conflicting findings. Numerous studies using a variety of behavioral paradigms
Statistical Regularities Shape Semantic Development 51
have provided evidence that taxonomic relations only begin to contribute at the group
level relatively late in development (e.g., Blaye et al., 2006; Lucariello et al., 1992;
Tversky, 1985; Walsh et al., 1993), and a similarly large body of studies have provided
evidence for early taxonomic organization (e.g., Bauer & Mandler, 1989; Deák & Bauer,
1996; Gelman & Markman, 1986; Waxman & Namy, 1997). In spite of the contradiction
between these bodies of research, we propose here that our present findings can be
reconciled with both.
Evidence for Late Taxonomic Onset. Results from several prior studies using a
variety of behavioral paradigms that have observed an influence of taxonomic relatedness
only in older children (e.g., age six and above), often following the earlier emergence of
influences of relations that may be derived from co-occurrence or perceptual similarity.
For example, many prior studies have investigated children’s semantic organization using
match-to-sample paradigms, in which participants are presented with a sample item (e.g.,
dog), and two choice items that are each related to the target in a different way (e.g.,
elephant and bone), and must select one choice item to match to the sample. Some
studies that have used this approach have observed that, although older children may
reliably choose taxonomic matches, young children do not (Lucariello et al., 1992;
Tversky, 1985; Walsh et al., 1993). A similar pattern in which a robust influence of
taxonomic relatedness is observed only in older children has emerged from studies that
have inferred knowledge of semantic relations from sorting (Blaye et al., 2006), list recall
(Bjorklund & Jacobs, 1985; Monnier & Bonthoux, 2011), and word association (Nelson,
1977)
5
paradigms. The present findings provide nuance to this apparent trajectory by
5
One exception to this pattern in word association patterns is the tendency for even young children to
produce taxonomic (or “paradigmatic”) responses to number words, such as responding “two” when
Statistical Regularities Shape Semantic Development 52
suggesting that an influence of taxonomic relations is not entirely absent in young
children, but is instead comparatively weak and uncommon, such that it is more readily
detected when using a sensitive, graded measure such as the dynamics of looking
behavior measured in Experiment 3.
Evidence for Early Taxonomic Onset. Oher prior studies have yielded results that
appear to demonstrate taxonomic knowledge that is detectable at the group level early in
development. Specifically, in studies using variants of the match-to-sample paradigm
conducted by Bauer and Mandler (1989); Deák and Bauer (1996); Gelman and Markman
(1986); and Waxman and Namy (1997), young children consistently chose taxonomic
matches, either throughout the study or under specific conditions. The evidence from our
present experiments also suggests contributions of taxonomic relations in young children,
and is only inconsistent with these prior findings in the strength and prevalence of these
contributions.
One potential explanation for this difference in strength and prevalence of
taxonomic knowledge is that additional information that could support taxonomic
choices was available in prior studies showing strong, prevalent taxonomic influences.
For example, in some of these 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, in some of these studies, targets and
taxonomic matches were given either identical labels, which may act as perceptual
features that contribute to similarity in young children (Sloutsky & Fisher, 2004), or co-
prompted with the word “one”. However, our analyses of co-occurrence in child-directed speech measured
from CHILDES corpora suggest that nouns for numbers one through ten frequently co-occur, rendering it
unclear whether these responses are driven by co-occurrence, or an understanding that number words
belong to the same category.
Statistical Regularities Shape Semantic Development 53
occurring labels (e.g., puppy and dog), such that taxonomic choices could be based on
co-occurrence (Fisher, 2010; Fisher, Matlen, & Godwin, 2011). Similarly, 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 (Bergelson
& Aslin, 2017; Styles & Plunkett, 2009; Willits, Wojcik, Seidenberg, & Saffran, 2013).
To the authors’ knowledge, the only group-level evidence of an influence of
taxonomic relatedness in young children in the absence of additional supportive
information comes from one of two cued-recall paradigm experiments conducted by
Blewitt and Toppino (1991). Specifically, Blewitt and Toppino found that recall accuracy
in preschool-age children given pairs of unrelated words was exceeded by the accuracy
of children given pairs of words that another sample of children had judged as co-
occurring in both experiments, but was only also exceeded by the accuracy of children
given taxonomically related words (referred to as “coordinate” pairs) in Experiment 2.
Although the authors identified the lack of a taxonomic influence on accuracy in
Experiment 1 as “spurious” (p. 311, Blewitt & Toppino, 1991), this inconsistency at least
suggests that the taxonomic influence was less robust and evident only in some children,
just as in the results of our experiments.
Finally, we note that the present findings rule out an alternative explanation that the
apparent weakness of taxonomic relations in children was simply due to the possibility
that the paradigms used in the present experiments were more sensitive to co-occurrence
than taxonomic relations. Specifically, such a bias in the paradigms would have also led
to the appearance of stronger co-occurrence versus taxonomic effects in adults. In
contrast, we observed similar co-occurrence and taxonomic effects in adults.
Statistical Regularities Shape Semantic Development 54
Taken together, although the evidence available from prior research is sufficiently
equivocal to fuel further debate, the evidence supporting the possibility that development
typically involves an early-emerging role for co-occurrence that is increasingly
supplemented by taxonomic relatedness is also sufficiently strong to highlight the
importance of developing accounts that can explain this trajectory. This topic is discussed
further in the following section.
Future Directions
The results of the present experiments were most consistent with the predictions of
the mechanistic account proposed by Sloutsky et al. (2017). Specifically, in this account,
sensitivity to co-occurrence regularities fosters the formation of semantic relations
between both concepts whose referents or labels directly co-occur with each other (e.g.,
fork and spaghetti) and concepts whose referents or labels share patterns of co-
occurrence (e.g., spaghetti and pie, which both co-occur with fork), which are in turn often
taxonomically related (Asr et al., 2016; Cree & Armstrong, 2012; Huebner & Willits, 2018;
Jones et al., 2015; Landauer & Dumais, 1997). According to this perspective, the earlier
influence on semantic knowledge of co-occurrence versus taxonomic relatedness occurs
because the latter is derived from the former. This prediction was corroborated by the
developmental trajectory observed in the present experiments.
However, the core mechanisms proposed in Sloutsky et al.’s (2017) account, in which
semantic relations between words are formed purely based on regularities with which
they either directly co-occur or share each other’s patterns of co-occurrence, remain
largely unexplored in human learners. Specifically, prior research investigating this
possibility is limited to only a handful of recent studies suggesting that toddlers and
Statistical Regularities Shape Semantic Development 55
children form relations between words that directly co-occur (Matlen, Fisher, & Godwin,
2015; Wojcik & Saffran, 2015). Therefore, one key future direction highlighted by the
present experiments is to test whether exposure to empirically manipulated linguistic input
in which some pairs of words directly co-occur, and others share each other’s patterns of
co-occurrence, does indeed drive the formation of corresponding semantic relations in
children and adults.
Conclusions
The present experiments provided evidence that word-word co-occurrence
regularities captures relations between concepts in the semantic organization of both
young children and adults. With development, co-occurrence was supplemented rather
than replaced by taxonomic relatedness. These findings highlight importance of
developing theoretical accounts of semantic development that incorporate a key role for
co-occurrence regularities from early childhood onward.
Statistical Regularities Shape Semantic Development 56
References
Princeton University. (2010). About WordNet. http://wordnet.princeton.edu
Arias-Trejo, N., & Plunkett, K. (2009). Lexical–semantic priming effects during infancy.
Philosophical Transactions of the Royal Society B: Biological Sciences, 364, 3633-3647.
Asr, F. T., Willits, J. A., & Jones, M. N. (2016). Comparing predictive and co-occurrence based
models of lexical semantics trained on child-directed speech Proceedings of the 38th
Annual Meeting of the Cognitive Science Society. Philadelphia, PA.
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, 390-412.
Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting Linear Mixed-Effects Models
Using lme4. Journal of Statistical Software, 67, 1-48.
Bauer, P. J., & Mandler, J. M. (1989). Taxonomies and triads: Conceptual organization in one-to
two-year-olds. Cognitive Psychology, 21, 156-184.
Bergelson, E., & Aslin, R. N. (2017). Nature and origins of the lexicon in 6-mo-olds. Proceedings
of the National Academy of Sciences, 114, 12916-12921.
Bjorklund, D. F., & Jacobs, J. W. (1985). Associative and categorical processes in children's
memory: The role of automaticity in the development of organization in free recall. Journal
of Experimental Child Psychology, 39, 599-617.
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.
Blewitt, P., & Toppino, T. C. (1991). The development of taxonomic structure in lexical memory.
Journal of Experimental Child Psychology, 51, 296-319.
Statistical Regularities Shape Semantic Development 57
Bower, G. H., Clark, M. C., Lesgold, A. M., & Winzenz, D. (1969). Hierarchical retrieval schemes
in recall of categorized word lists. Journal of Verbal Learning and Verbal Behavior, 8, 323-
343.
Bulf, H., Johnson, S. P., & Valenza, E. (2011). Visual statistical learning in the newborn infant.
Cognition, 121, 127-132.
Carey, S. (1985). Conceptual change in childhood. Cambridge, Massachusetts: MIT Press.
Cree, G. S., & Armstrong, B. C. (2012). Computational models of semantic memory The
Cambridge Handbook of Psycholinguistics (pp. 259-282). Cambridge: Cambridge
University Press.
Deák, G., & Bauer, P. (1996). The dynamics of preschoolers' categorization choices. Child
Development, 67, 740-767.
Fenson, L., Vella, D., & Kennedy, M. (1989). Children's knowledge of thematic and taxonomic
relations at two years of age. Child Development, 60, 911-919.
Fisher, A. V. (2010). What’s in the name? Or how rocks and stones are different from bunnies
and rabbits. Journal of Experimental Child Psychology, 105, 198-212.
Fisher, A. V., Matlen, B. J., & Godwin, K. E. (2011). Semantic similarity of labels and inductive
generalization: Taking a second look. Cognition, 118, 432-438.
Fox, J., & Weisberg, S. (2011). An R Companion to Applied Regression (Second Edition ed.).
Thousand Oaks, CA: Sage.
Frank, M. C., Braginsky, M., Yurovsky, D., & Marchman, V. A. (2016). Wordbank: An open
repository for developmental vocabulary data. Journal of Child Language.
Frermann, L., & Lapata, M. (2015). Incremental Bayesian Category Learning From Natural
Language. Cognitive Science, 40, 1333–1381.
Fulkerson, A. L., & Waxman, S. R. (2007). Words (but not tones) facilitate object categorization:
Evidence from 6-and 12-month-olds. Cognition, 105, 218-228.
Statistical Regularities Shape Semantic Development 58
Gellatly, A. R., & Gregg, V. H. (1975). The effects of negative relatedness upon wordpicture and
wordword comparisons and subsequent recall. British Journal of Psychology, 66, 311-
323.
Gelman, S. A., & Coley, J. D. (1990). The importance of knowing a dodo is a bird: Categories and
inferences in 2-year-old children. Developmental Psychology, 26, 796.
Gelman, S. A., & Markman, E. M. (1986). Categories and induction in young children. Cognition,
23, 183-209.
Goldberg, R. F., & Thompson-Schill, S. L. (2009). Developmental “roots” in mature biological
knowledge. Psychological Science, 20, 480-487.
Heit, E. (2000). Properties of inductive reasoning. Psychonomic Bulletin & Review, 7, 569-592.
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.
Howard, D. V., & Howard, J. H. (1977). A multidimensional scaling analysis of the development
of animal names. Developmental Psychology, 13, 108.
Huebner, P. A., & Willits, J. A. (2018). Structured semantic knowledge can emerge automatically
from predicting word sequences in child-directed speech. Frontiers in Psychology, 9.
Huettig, F., & Altmann, G. T. (2005). Word meaning and the control of eye fixation: Semantic
competitor effects and the visual world paradigm. Cognition, 96, B23-B32.
Huettig, F., Quinlan, P. T., McDonald, S. A., & Altmann, G. T. (2006). Models of high-dimensional
semantic space predict language-mediated eye movements in the visual world. Acta
Psychologica, 121, 65-80.
Inhelder, B., & Piaget, J. (1964). The early growth of logic in the child. New York: Norton.
Jimura, K., Hirose, S., Wada, H., Yoshizawa, Y., Imai, Y., Akahane, M., . . . Konishi, S. (2016).
Relatedness-dependent rapid development of brain activity in anterior temporal cortex
during pair-association retrieval. Neuroscience Letters, 627, 24-29.
Statistical Regularities Shape Semantic Development 59
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.
Kemp, C., & Tenenbaum, J. B. (2008). The discovery of structural form. Proceedings of the
National Academy of Sciences, 105, 10687-10692.
Klem, M., MelbyLervåg, M., Hagtvet, B., Lyster, S. A. H., Gustafsson, J. E., & Hulme, C. (2015).
Sentence repetition is a measure of children's language skills rather than working memory
limitations. Developmental Science, 18, 146-154.
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, 978-998.
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.
McClelland, J. L., & Rogers, T. T. (2003). The parallel distributed processing approach to
semantic cognition. Nature Reviews Neuroscience, 4, 310-322.
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.
Mirman, D., & Graziano, K. M. (2012). Individual differences in the strength of taxonomic versus
thematic relations. Journal of Experimental Psychology: General, 141, 601.
Statistical Regularities Shape Semantic Development 60
Mirman, D., & Magnuson, J. S. (2009). Dynamics of activation of semantically similar concepts
during spoken word recognition. Memory & Cognition, 37, 1026-1039.
Monnier, C., & Bonthoux, F. (2011). The semanticsimilarity effect in children: Influence of long
term knowledge on verbal shortterm memory. British Journal of Developmental
Psychology, 29, 929-941.
Murphy, G. L. (2001). Causes of taxonomic sorting by adults: A test of the thematic-to-taxonomic
shift. Psychonomic Bulletin & Review, 8, 834-839.
Nelson, K. (1977). The syntagmatic-paradigmatic shift revisited: a review of research and theory.
Psychological Bulletin, 84, 93.
Rosch, E. (1975). Basic objects in natural categories: Language Behavior Research Laboratory,
University of California.
Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B. Lloyd (Eds.), Cognition and
Categorization (pp. 27-48): Hillsdale, NJ: Lawrence Erbaum Associates.
Ross, B. H., & Murphy, G. L. (1999). Food for thought: Cross-classification and category
organization in a complex real-world domain. Cognitive Psychology, 38, 495-553.
Sadeghi, Z., McClelland, J. L., & Hoffman, P. (2015). You shall know an object by the company it
keeps: An investigation of semantic representations derived from object co-occurrence in
visual scenes. Neuropsychologia, 76, 52-61.
Samuelson, L. K., & Smith, L. B. (1999). Early noun vocabularies: do ontology, category structure
and syntax correspond? Cognition, 73, 1-33.
Sloutsky, V. M. (2010). From perceptual categories to concepts: What develops? Cognitive
Science, 34, 1244-1286.
Sloutsky, V. M., & Fisher, A. V. (2004). Induction and categorization in young children: a similarity-
based model. Journal of Experimental Psychology: General, 133, 166-187.
Sloutsky, V. M., Yim, H., Yao, X., & Dennis, S. (2017). An associative account of the development
of word learning. Cognitive Psychology, 97, 1-30.
Statistical Regularities Shape Semantic Development 61
Smiley, S. S., & Brown, A. L. (1979). Conceptual preference for thematic or taxonomic relations:
A nonmonotonic age trend from preschool to old age. Journal of Experimental Child
Psychology, 28, 249-257.
Smith, L. B., & Heise, D. (1992). Perceptual similarity and conceptual structure. In B. Burns (Ed.),
Advances in Psychology: Percepts, Concepts, and Categories (Vol. 93, pp. 233-272).
North-Holland: Elsevier.
Spence, D. P., & Owens, K. C. (1990). Lexical co-occurrence and association strength. Journal
of Psycholinguistic Research, 19, 317-330.
Storm, C. (1980). The semantic structure of animal terms: A developmental study. International
Journal of Behavioral Development, 3, 381-407.
Styles, S. J., & Plunkett, K. (2009). How do infants build a semantic system? Language and
Cognition, 1, 1-24.
Tversky, B. (1985). Development of taxonomic organization of named and pictured categories.
Developmental Psychology, 21, 1111-1119.
Walsh, M., Richardson, K., & Faulkner, D. (1993). Perceptual, thematic and taxonomic relations
in children’s mental representations: Responses to triads. European Journal of
Psychology of Education, 8, 85-102.
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.
Wisniewski, E. J., & Bassok, M. (1999). What makes a man similar to a tie? Stimulus compatibility
with comparison and integration. Cognitive Psychology, 39, 208-238.
Wojcik, E. H., & Saffran, J. R. (2015). Toddlers encode similarities among novel words from
meaningful sentences. Cognition, 138, 10-20.
Statistical Regularities Shape Semantic Development 62
Appendix A
Table A1
T-scores and Resnik similarities for pairs in the Co-Occur,
Taxonomic, and Unrelated conditions in Experiments 1 & 2.
Item 1
Item 2
t-score
Resnik
Co-Occur
bottle
baby
6.13
1.37
foot
shoe
5.30
0.61
brush
hair
15.75
1.37
cup
juice
7.68
0.61
cheese
mouse
2.97
0.61
car
street
4.40
2.49
soup
spoon
3.09
0.61
milk
cow
4.53
0.61
paper
pencil
6.61
0.61
Taxonomic
ball
puzzle
-3.65
7.98
pig
bear
-1.91
5.61
horse
bunny
-0.54
5.61
carrot
banana
-0.19
6.35
fork
bowl
0.21
8.10
popcorn
fries
0.00
6.97
airplane
boat
0.83
6.47
sock
pajamas
0.00
5.87
chicken
owl
-1.38
6.94
Unrelated
crayon
frog
0.00
1.37
towel
bread
0.12
0.61
blocks
cereal
0.00
0.61
balloon
tree
0.00
1.37
sheep
pancake
0.00
0.61
pizza
lion
0.96
0.61
fish
bed
-8.73
1.37
duck
swing
0.00
1.37
ice cream
bicycle
0.00
0.61
Note. T-scores for word pairs that never co-occurred within the
7-word window are undefined. Values for these pairs have
therefore been entered as 0.00.
Appendix B
Table B1
T-scores and Resnik similarities for pairs in the Co-Occur, Taxonomic, and Unrelated conditions in Experiment 3.
Target
Co-Occur
t-score
Resnik
Taxonomic
t-score
Resnik
Unrelated
t-score
Resnik
Car
Street
4.40
2.49
Bicycle
0.71
6.31
Stick
-2.68
1.37
Nose Tissue 26.33 0.61 Tongue 0.34 5.21 0.18 0.61
Airplane
Sky
3.94
0.61
Boat
0.83
6.47
Bathtub
0.00
3.45
Cheese Mouse 2.97 0.61 Ice Cream 0.00 5.46 0.00 0.61
Horse
Cowboy
2.09
1.82
Frog
-0.11
5.20
Cloud
0.00
0.61
Pizza Oven 6.38 0.61 Chocolate 0.00 5.46 0.00 0.61
Bread
Knife
3.89
0.61
Muffin
0.48
8.51
Button
0.04
0.61
Foot Shoe 5.30 0.61 Head -0.20 4.65 0.00 0.61
Fish
Water
5.65
0.61
Bird
1.17
5.20
Train
-1.29
1.37
Bed Pillow 6.19 2.49 Table -0.43 6.19 -1.62 3.45
Cow
Milk
4.53
0.61
Tiger
0.44
5.61
Book
-1.58
1.37
Leg Pants 2.06 0.61 Finger 0.28 6.06 -2.94 0.61
Cereal
Breakfast
5.17
3.45
Pancake
0.00
6.58
Shovel
0.00
0.61
Monkey Zoo 3.67 1.37 Squirrel 0.00 5.61 0.00 1.37
Soup
Spoon
3.09
0.61
Juice
0.01
5.46
Telephone
0.00
0.61
Coat Zipper 2.75 2.49 Sweater 1.45 6.78 -0.70 2.49
Bottle
Baby
6.13
1.37
Bowl
0.00
6.95
Door
0.00
2.49
Apple Tree 3.16 1.37 Grapes 0.00 8.00 -2.21 1.37
Corn
Chicken
3.22
0.61
Pumpkin
0.34
5.91
Star
0.00
0.61
Sock Foot 4.93 0.61 Hat 0.54 5.87 0.00 1.37
Owl
Moon
5.25
1.37
Turtle
0.00
5.20
Nail
0.00
1.37
Bucket
Hose
5.99
2.49
Jar
0.45
5.32
0.00
3.45
Note. T-scores for word pairs that never co-occurred within the 7-word window are undefined. Values for these pairs have
therefore been entered as 0.00
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Previous research has suggested that distributional learning mechanisms may contribute to the acquisition of semantic knowledge. However, distributional learning mechanisms, statistical learning, and contemporary “deep learning” approaches have been criticized for being incapable of learning the kind of abstract and structured knowledge that many think is required for acquisition of semantic knowledge. In this paper, we show that recurrent neural networks, trained on noisy naturalistic speech to children, do in fact learn what appears to be abstract and structured knowledge. We trained two types of recurrent neural networks (Simple Recurrent Network, and Long Short-Term Memory) to predict word sequences in a 5-million-word corpus of speech directed to children ages 0–3 years old, and assessed what semantic knowledge they acquired. We found that learned internal representations are encoding various abstract grammatical and semantic features that are useful for predicting word sequences. Assessing the organization of semantic knowledge in terms of the similarity structure, we found evidence of emergent categorical and hierarchical structure in both models. We found that the Long Short-term Memory (LSTM) and SRN are both learning very similar kinds of representations, but the LSTM achieved higher levels of performance on a quantitative evaluation. We also trained a non-recurrent neural network, Skip-gram, on the same input to compare our results to the state-of-the-art in machine learning. We found that Skip-gram achieves relatively similar performance to the LSTM, but is representing words more in terms of thematic compared to taxonomic relations, and we provide reasons why this might be the case. Our findings show that a learning system that derives abstract, distributed representations for the purpose of predicting sequential dependencies in naturalistic language may provide insight into emergence of many properties of the developing semantic system.
Article
Full-text available
The MacArthur-Bates Communicative Development Inventories (CDIs) are a widely used family of parent-report instruments for easy and inexpensive data-gathering about early language acquisition. CDI data have been used to explore a variety of theoretically important topics, but, with few exceptions, researchers have had to rely on data collected in their own lab. In this paper, we remedy this issue by presenting Wordbank, a structured database of CDI data combined with a browsable web interface. Wordbank archives CDI data across languages and labs, providing a resource for researchers interested in early language, as well as a platform for novel analyses. The site allows interactive exploration of patterns of vocabulary growth at the level of both individual children and particular words. We also introduce wordbankr , a software package for connecting to the database directly. Together, these tools extend the abilities of students and researchers to explore quantitative trends in vocabulary development.
Article
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.
Article
Full-text available
This study examined 2-year-old childrern's ability to make category-based inferences. Subjects were asked a series of questions that they could answer based on category membership, appearances, or both. In one condition, all pictures were named; in a second condition, none were named. Children performed well on prototypical pictures regardless of whether they were named; on atypical pictures, they performed better when category labels were provided. A control study demonstrated that children ignored the label when it named a transient property rather than a stable category. Contrary to standard views of young children, these results indicate an early-emerging capacity to overlook salient appearances. However, one important development still to take place is the ability to use subtle perceptual cues to determine category membership in the absence of language.
Article
Full-text available
An influential position in lexical semantics holds that semantic representations for words can be derived through analysis of patterns of lexical co-occurrence in large language corpora. Frith (1957) famously summarised this principle as "you shall know a word by the company it keeps". We explored whether the same principle could be applied to non-verbal patterns of object co-occurrence in natural scenes. We performed latent semantic analysis (LSA) on a set of photographed scenes in which all of the objects present had been manually labelled. This resulted in a representation of objects in a high-dimensional space in which similarity between two objects indicated the degree to which they appeared in similar scenes. These representations revealed similarities among objects belonging to the same taxonomic category (e.g., vehicles, items of clothing) as well as cross-category associations (e.g., between fruits and kitchen utensils). We also compared representations generated from this scene dataset with two established methods for elucidating semantic representations: (a) a published database of semantic features generated verbally by participants and (b) LSA applied to a linguistic corpus in the usual fashion. Statistical comparisons of the three methods indicated significant association between the structures revealed by each method, with the scene dataset displaying greater convergence with feature-based representations than did LSA applied to linguistic data. The results indicate that information about the conceptual significance of objects can be extracted from their patterns of co-occurrence in natural environments, opening the possibility for such data to be incorporated into existing models of conceptual representation.
Chapter
Full-text available
Meaning is a fundamental component of nearly all aspects of human cognition, but formal models of semantic memory have classically lagged behind many other areas of cognition. However, computational models of semantic memory have seen a surge progress in the last two decades, advancing our knowledge of how meaning is constructed from experience, how knowledge is represented and used, and what processes are likely to be culprit in disorders characterized by semantic impairment. This chapter provides an overview of several recent clusters of models and trends in the literature, including modern connectionist and distributional models of semantic memory, and contemporary advances in grounding semantic models with perceptual information and models of compositional semantics. Several common lessons have emerged from both the connectionist and distributional literatures, and we attempt to synthesize these themes to better focus future developments in semantic modeling.
Article
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.
Article
Models of category learning have been extensively studied in cognitive science and primarily tested on perceptual abstractions or artificial stimuli. In this paper, we focus on categories acquired from natural language stimuli, that is, words (e.g., chair is a member of the furniture category). We present a Bayesian model that, unlike previous work, learns both categories and their features in a single process. We model category induction as two interrelated subproblems: (a) the acquisition of features that discriminate among categories, and (b) the grouping of concepts into categories based on those features. Our model learns categories incrementally using particle filters, a sequential Monte Carlo method commonly used for approximate probabilistic inference that sequentially integrates newly observed data and can be viewed as a plausible mechanism for human learning. Experimental results show that our incremental learner obtains meaningful categories which yield a closer fit to behavioral data compared to related models while at the same time acquiring features which characterize the learned categories. (An earlier version of this work was published in Frermann and Lapata .).
Article
In two experiments subjects were required to compare the meanings of either a word and a picture, or of two words. Different name levels, i.e. category versus superordinate names, had only a small effect on the time to compare a name with a picture. When incongruent stimulus pairs were semantically related, both positive and negative decision times were longer than when the incongruent pairs were unrelated; relatedness also affected subjects' recall of stimuli. Implications for models of semantic decision are discussed.