ArticlePDF AvailableLiterature Review

Using Network Science to Understand the Aging Lexicon: Linking Individuals' Experience, Semantic Networks, and Cognitive Performance



People undergo many idiosyncratic experiences throughout their lives that may contribute to individual differences in the size and structure of their knowledge representations. Ultimately, these can have important implications for individuals' cognitive performance. We review evidence that suggests a relationship between individual experiences, the size and structure of semantic representations, as well as individual and age differences in cognitive performance. We conclude that the extent to which experience-dependent changes in semantic representations contribute to individual differences in cognitive aging remains unclear. To help fill this gap, we outline an empirical agenda that utilizes network analysis and involves the concurrent assessment of large-scale semantic networks and cognitive performance in younger and older adults. We present preliminary data to establish the feasibility and limitations of such empirical, network-analytical approaches.
Topics in Cognitive Science 14 (2022) 93–110
© 2021 The Authors. Topics in Cognitive Science published by Wiley Periodicals LLC on behalf of Cognitive
Science Society
ISSN: 1756-8765 online
DOI: 10.1111/tops.12586
This article is part of the topic “Networks of the Mind: How Can Network Science Elucidate
Our Understanding of Cognition?,” Thomas T. Hills and Yoed N. Kenett (Topic Editors).
Using Network Science to Understand the Aging Lexicon:
Linking Individuals’ Experience, Semantic Networks, and
Cognitive Performance
Dirk U. Wulff,a,b Simon De Deyne,cSamuel Aeschbach,a
Rui Mataa,b
aFaculty of Psychology, University of Basel
bCenter for Adaptive Rationality, Max Planck Institute for Human Development
cMelbourne School of Psychological Sciences, University of Melbourne
Received 15 February 2021; received in revised form 19 October 2021; accepted 19 October 2021
People undergo many idiosyncratic experiences throughout their lives that may contribute to indi-
vidual differences in the size and structure of their knowledge representations. Ultimately, these can
have important implications for individuals’ cognitive performance. We review evidence that suggests
a relationship between individual experiences, the size and structure of semantic representations, as
well as individual and age differences in cognitive performance. We conclude that the extent to which
experience-dependent changes in semantic representations contribute to individual differences in cog-
nitive aging remains unclear. To help fill this gap, we outline an empirical agenda that utilizes net-
work analysis and involves the concurrent assessment of large-scale semantic networks and cognitive
Correspondence should be sent to Dirk U. Wulff, Faculty of Psychology, Cognitive and Decision Sciences,
University of Basel, Missionsstrasse 60-62, 4055 Basel, Switzerland. E-mail:
This is an open access article under the terms of the Creative Commons Attribution License, which permits
use, distribution and reproduction in any medium, provided the original work is properly cited.
[Correction added on 21st April 2022, after first online publication: CSAL funding statement has been added.]
94 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
performance in younger and older adults. We present preliminary data to establish the feasibility and
limitations of such empirical, network-analytical approaches.
Keywords: Semantic networks; Cognitive aging; Individual differences
1. Introduction
From childhood and adolescence onwards, the average human reads a couple of books each
year, watches hundreds of hours of TV, and spends many hours on social media (Twenge,
Martin, & Spitzberg, 2019), leading to the accumulation of a large and, potentially, largely
unique, set of experiences during a lifetime. To what extent does the accumulation of such
idiosyncratic experiences contribute to individual differences in thought and judgment across
the life span?
Aging research has long realized not only the importance of describing the modal changes
in cognition across the life span but also that “cognitive development in adulthood and old
age differs substantially from person to person and is malleable within individuals” (Linden-
berger, 2014 p. 576). Despite the field’s direct acknowledgment of interindividual differences,
we still know little about the sources of such differences and the extent to which idiosyncratic
life experiences contribute to cognitive performance. Crucially, some voices have raised the
possibility that cumulative experience and the cognitive representations they give rise to are a
major factor underlying typical age-related patterns, such as decreased memory performance
with increased age (Buchler & Reder, 2007; Ramscar, Hendrix, Shaoul, Milin, & Baayen,
In this article, we propose that network science can be instrumental in illuminating the
effects of cumulative experience on individual differences in cognition. To this end, we first
review the existing literature on the links between cumulative experience, cognitive repre-
sentation, and cognitive performance. We then identify the lack of direct assessment of indi-
viduals’ mental representations as a central limitation, and propose an empirical agenda that
utilizes network analysis to fill this gap. Next, we present a proof-of-concept study that illus-
trates the feasibility and limitations of our approach involving the elicitation of large-scale
semantic networks from single individuals. Finally, we discuss the challenges and implica-
tions of the outlined research agenda to understand individual and age differences.
2. From experience to cognitive performance: An overview of the current literature
In this section, we assess the current state of the psychological literature that has investi-
gated the links between environmental exposure, the size and structure of semantic represen-
tations, and cognitive performance across the life span (see Fig. 1). Our goal is both to provide
a brief overview of the current literature and to highlight the main gaps that must be filled to
understand the role of individual experience in cognitive representations and performance.
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 95
Fig. 1. Research pathways that have linked environment, representation, and cognitive performance across the
life span. Pathway (A) represents research studying the association between the environment and cognitive perfor-
mance without explicit consideration of the underlying mental representations. Pathway (B) represents research on
the association between environmental exposure and mental representations. Pathway (C) represents research on
the association between mental representations and cognitive performance. We propose that pathways (B) and (C)
can particularly benefit from a network analytic approach.
There are several lines of research supporting a close link between cumulative experience
and cognitive performance (see Fig. 1, pathway A: Environment–Performance), of which
three stand out. First, research that summarizes the association between print exposure and
reading performance in children and young adults underlines the tremendous impact of expe-
rience on linguistic proficiency. Specifically, meta-analytic results show moderate to strong
correlations between print exposure and reading comprehension or spelling (Mol & Bus,
2011). This work supports the idea of an “upward spiral,” such that individuals who are more
exposed to written text become more proficient in reading and comprehension, which, in turn,
leads to increased print exposure, further increasing linguistic proficiency. Second, one of the
best documented findings in aging research concerns the observed increases in crystallized
abilities across the life span, as reflected in vocabulary size (Verhaeghen, 2003, for a meta-
analysis). This work suggests that even as adults, individuals continue to expand their vocab-
ulary store throughout their lives (Brysbaert, Stevens, Mandera, & Keuleers, 2016). This is
compatible with the idea that such vocabulary growth is related to individuals’ lifetime formal
or informal education. Notably, such increases in knowledge seem to have general implica-
tions for cognitive performance. For example, older adults are more likely to rely on predic-
tion in reading because of their additional reading experience (Huettig & Pickering, 2019; but
see Wlotko, Federmeier, & Kutas, 2012). Third, research on expertise suggests that increased
cumulative experience leads to domain-specific memory performance (Sala & Gobet, 2017,
for a meta-analysis), such as experienced chess players being able to memorize both realis-
tic and random chess positions better than novices (Gobet & Simon, 1996). All in all, such
96 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
findings coalesce to make a strong case for the importance of environmental effects on cog-
nitive performance and suggest that these effects are cumulative and, likely, domain specific.
Most, if not all, psychologists will find it trivial to state that the effects of cumulative expe-
rience are somehow mediated by its effects on mental representations and processes (Fig. 1,
pathway B, Environment–Representation). Despite the truism and the increasing consen-
sus that the lexical-semantic space continues to be shaped by personal linguistic experience
throughout the life span (Rodd, 2020), only recently have researchers started to probe more
deeply into the effects of cumulative experience on the structure of knowledge representa-
tions. Some researchers have adopted graph-based approaches to capture potential structural
changes to the mental lexicon that occur across the life span as a function of experience (for
an overview, see Wulff, De Deyne, Jones, Mata, & The Aging Lexicon Consortium, 2019).
The current literature suggests that cumulative experience has implications for not only the
size but also the structure of mental representations (e.g., Cosgrove, Kenett, Beaty, & Diaz,
2021; Dubossarsky, De Deyne, & Hills, 2017). For example, Dubossarsky and colleagues
conducted a network analysis of free association data from thousands of individuals (10 to 84
years of age), and found older adults’ semantic networks were less connected (i.e., the words
in the network have lower average degrees), less organized (i.e., the words in the network
have a lower average local clustering coefficient), and less efficient (i.e., the shortest path
length between any two words in the network is greater) relative to those of younger adults.
Crucially, new methods are becoming available that promise ever greater insight and preci-
sion in mapping individual and group differences in the structure of mental representations
(e.g., Benedek et al., 2017; Morais, Olsson, & Schooler, 2013; Zemla & Austerweil, 2018).
Table 1 presents an overview of a few approaches to measuring individual-level mental rep-
resentations and we discuss in more detail pros and cons of each approach in the next section
(cf. Section 3).
Finally, research is accumulating that establishes direct links between the structure of
semantic representations and cognitive performance (Fig. 1, pathway C, Representation–
Performance). For example, various studies on memory recall show impaired performance
when words represented as nodes within a semantic network have lower clustering (Nelson,
Bennett, Gee, Schreiber, & McKinney, 1993). More recent work has expanded this line of
research to understand how lexical and semantic structure is crucial to individual cognitive
performance in various domains, such as intelligence and creativity (He et al., 2021; Kenett
& Faust, 2019). These more recent studies are particularly relevant because they have started
relying on individual or small group estimates of mental representations and how macroprop-
erties of representational networks impact cognitive performance. However, one should note
that this work has not established a direct link between explicit measures of an individual’s
cumulative experience, such as books read or movies watched, and the size or structure of the
mental representations—instead studies have typically relied on proxies for experience, such
as age (Wulff et al., 2019).
All in all, these different lines of research support the idea that cumulative personal expe-
rience plays a crucial role in determining individual differences in cognitive performance
across the life span. However, much of this work is correlational (Nation, 2017, for a sim-
ilar critique) and not carried out at the level of the individual. As a result, the mechanisms
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 97
Tab l e 1
Approaches to measuring individual-level semantic networks
Paradigm Description Scope Comparability References
Verbal fluency Individuals generate as many items from
a given category as they can in a fixed
period of time
Limited High (Zemla & Austerweil,
Individuals rate the relation, i.e.,
similarity, between pairs of items
Limited High (Benedek et al., 2017;
Roads & Love, 2020)
Free association
Individuals generate one or more
associations to word cues, which are
participant generated
Broad Low (Morais et al., 2013)
Free association
(fixed list)
Individuals generate one or more
associations to experimenter-generated
word cues
Broad High See below
Note.Scope =Ability of the paradigm to provide coverage of a large set of semantic categories; Comparability =Ability of the paradigm to provide
comparable coverage of semantic representation from different individuals.
98 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
that tie cumulative exposure to performance are still under investigation. We propose that
to truly understand the consequences of individual experience on mental representation and
cognitive performance, the concurrent assessment of individuals’ unique environments, as
well as their mental representations and cognitive performance, is needed. In this endeavor,
network science provides a key tool to capturing individuals’ mental representations, in par-
ticular, in the quantification of the connection patterns between mental concepts that can be
empirically related to both experience and cognitive performance. In what follows, we offer
an empirical agenda that uses network science tools to better understand the aging lexicon.
3. An empirical agenda
As outlined above, the current literature has made large strides toward understanding the
components linking cumulative experience to cognitive performance: We know that younger
and older adults differ in the amounts and kinds of experiences, the contents and structure of
mental representations, and that there are systematic age and individual differences in both
fluid and crystallized performance. However, presently, we cannot confidently estimate the
portion of individual differences in performance across the life span due to the accumulation
of specific types of experience. What is missing, in our mind, is a concerted empirical agenda
that investigates the Environment–Representation and the Representation–Performance path-
ways at the level of the individual, thus allowing us to quantify how a certain type or quantity
of experience translates into the size or structure of cognitive representations, and, ultimately,
into differences in cognitive performance. This goal presents challenges on several levels, and
we discuss two major ones below: first, the challenge of measuring individuals’ environments,
and, second, the challenge of capturing the content and structure of single individuals’ mental
Concerning the challenge of measuring individuals’ environments, one main limiting factor
is currently the lack of available data. Put simply, the field lacks context-aware longitudinal
projects that provide a characterization of the environments experienced by individuals over
time. Although unprecedented, large amounts of contextualized text and speech data are now
available to scientists (e.g., Love, Dembry, Hardie, Brezina, & McEnery, 2017; Schröter &
Schroeder, 2017), few of these data sets present data on the level of the individual or distin-
guish between age groups. Crucially, any individual’s environment consists of not only lin-
guistic information, but also rich multimodal sensorial information (e.g., De Deyne, Navarro,
Collell, & Perfors, 2021). Thus, a major challenge for future research is to create individual-
annotated, multimodal corpora accessible for research. There is optimism in the field that the
rise of the quantified-self movement and the availability of new “digital tracing” methods can
provide multiple data streams to feed computational modeling efforts that use these data to
create models of individual’s mental representations (Wulff et al., 2019). These models would
be crucial to generate expectations about individuals’ actual mental representations.
Concerning the challenge of mapping individuals’ mental representations, and as dis-
cussed above, there are now different approaches to obtaining semantic networks in order to
characterize the size and structure of mental representations (see Table 1). However, not all
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 99
of these approaches are equally suited to accurately uncovering the full breadth and depth
of a single individual’s mental representation (scope) while ensuring comparability between
individuals in the lexical-semantic space covered (comparability). For example, approaches
to derive networks derived from verbal fluency tasks, for example, using the U-INVITE algo-
rithm (Zemla & Austerweil, 2018), are limited to specific categories (e.g., concrete categories
such as animals) and semantic relationships (e.g. is-a). which may not be representative of
the entire semantic network (De Deyne & Storms, 2008). In addition, learning about within-
category structure is not very informative about the links between categories. In turn, other
approaches that derive networks from relatedness judgments require no intricate modeling
and can be powerful in providing comparability between individuals because participants can
be presented with the same pairs of concepts (Wulff, Hills, & Mata, 2018). However, asking
individuals to provide ratings to all possible pairs becomes prohibitively demanding with even
small sets of stimuli because this implies thousands of paired comparisons.1Compared to
other methods, free associations are a relatively economical basis for semantic networks and
can provide broad scope. However, free-association methods that ask individuals to generate
associations in a snowball method (Morais et al., 2013) may reduce comparability across par-
ticipants’ networks due to path dependency in any individual’s search of the representational
store. In comparison, to generate networks from free associations with fixed, experimenter-
generated lists may provide a better choice in both scope and comparability. One should note
that free association has also been the method of choice for most previous large-scale assess-
ments of aggregate semantic networks (De Deyne, Navarro, Perfors, Brysbaert, & Storms,
2019; Steyvers, Shiffrin, & Nelson, 2005).
All in all, our review of the literature, as well as assessment of the challenges above, high-
lights several future directions: first, providing a better description of the idiosyncratic expe-
riences of individuals that can inform computational and learning models of linguistic and
semantic cognition, and second, mapping the mental representations of single individuals
that can later be matched to expectations about the role of experience in cognitive perfor-
mance and the potential consequences. In what follows, we provide our own attempt at using
network science to tackle the second challenge of mapping individual semantic networks and
linking them to cognitive performance.
4. Estimating the Representation–Performance pathway: The MySWOW project
Our team is currently pursuing the My Small World of Words (MySWOW) project that
aims to map the semantic networks of single individuals so as to be able to link the structural
characteristics of each individual’s mental representation and several aspects of cumulative
experience and performance.
In what follows, we describe a small study that showcases and assesses the feasibility
of such an approach. Our project and the procedure of our feasibility study was inspired
by ongoing efforts to obtain word association norms for several languages in a large
online citizen-science project, the Small World of Words (SWOW) study (https://small SWOW has already offered a set of useful linguistic resources for both
100 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
Dutch (De Deyne, Navarro, & Storms, 2013) and English (De Deyne et al., 2019) and
aims to contribute additional resources concerning 15 other languages, including German
(, in the future. The adoption of the same procedure for the
large, population-based SWOW and the study of individuals, MySWOW, promises future
assessments of the comparability between the results from aggregate networks and those of
single individuals or small groups.
We largely adopted the procedure of SWOW, which asks participants to provide three
associates to a given cue (e.g., “cat”). The two additional responses help elicit nondominant
associations, with limited evidence of response chaining (i.e., bias from previous responses;
De Deyne et al., 2019). SWOW typically presents volunteers with 18 cues that provide
responses in the course of minutes. In contrast, we asked each individual participant to pro-
vide answers to thousands of cues over the course of weeks. Specifically, each participant in
our study was asked to provide three associations to a total of 3,000 unique cues and 600
repeated cues (to assess reliability), resulting in a total of 10,800 responses per participant.
We obtained data for four younger (aged 24 to 28) and four older (aged 68 to 70) native Ger-
man speakers. In an initial session, participants were briefed in person and received instruc-
tions concerning the dedicated online tool that they could use to complete the word asso-
ciations. Participants then completed the word associations from home over the course of
weeks. After finishing the word association task, participants returned to the laboratory for
cognitive testing on several memory and linguistic tasks. In particular, we focused on verbal
fluency, paired associated learning, and episodic memory tasks that have often been used to
estimate and understand individual and age differences in cognitive performance (Ramscar,
Sun, Hendrix, & Baayen, 2017; Zemla & Austerweil, 2018). A companion data paper pro-
vides a detailed account of our design, participants, and methods (Wulff, Aeschbach, Deyne,
& Mata, in press). The data can be downloaded from
4.1. Differences in the semantic networks of younger and older individuals
In what follows, we document a first effort to obtain individual semantic networks from
word association data in MySWOW and report a qualitative comparison of the results with
those from past work, which suggests structural differences between the semantic representa-
tions of younger and older adults (Dubossarsky et al., 2017; Wulff et al., 2019).
For each individual, we created unipartite networks by placing weighted, undirected edges
between responses and corresponding cues. This resulted in eight individual networks con-
taining 4,836 to 6,461 nodes, which is comparable in size to previously obtained large-scale
aggregate associative networks (Morais et al., 2013; Steyvers et al., 2005). Fig. 2 illustrates
the network of one individual, highlighting the most central words according to PageRank
(e.g., Griffiths, Steyvers, & Firl, 2007) and the underlying structure by identifying clusters
extracted using the Louvain method (Blondel et al., 2008).
Networks strongly overlapped in content with each other and with existing SWOW data
sets. For instance, the top 10 most central words in the German SWOW data set,2money,
music,work,school,water,car,love,green,andimportant, are all found among the top 100
most central words in all individual networks, while also making up 38% of the top 10 most
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 101
Fig. 2. Associative network of one participant. Node size represents PageRank centrality, colors communities
detected using the Louvain algorithm (Blondel, Guillaume, Lambiotte, & Lefebvre, 2008). Words were translated
from the original German and show the most important words in the respective community, with size representing
the word’s importance in the overall network as determined by PageRank.
Tab l e 2
Macroscropic differences in semantic network structure in terms of degree (k), clustering coefficient (C), and
average path length (L) for full and common individual networks based on nodes shared among participants
Full network Common network
Group Age |V|kCL|V|kCL
Young 24 5,780 3.03 .091 8.15 2,111 3.03 .153 7.66
Young 27 4,836 3.47 .115 7.13 2,111 3.55 .174 6.70
Young 27 4,920 3.31 .119 7.36 2,111 3.32 .202 6.81
Young 28 4,995 3.44 .136 7.14 2,111 3.62 .220 6.54
Old 68 5,275 3.35 .059 6.54 2,111 3.33 .110 5.94
Old 68 6,461 2.63 .045 9.33 2,111 2.46 .072 8.86
Old 69 6,157 2.78 .053 8.37 2,111 2.76 .093 7.76
Old 70 5,792 2.39 .055 11.3 2,111 2.19 .114 10.8
central words in individual networks. Consistent with previous work (e.g., Dubossarsky et al.,
2017; Wulff, Hills, Lachman, & Mata, 2016; Wulff et al., 2018), older adults’ networks were
larger (NOA =5,921 vs. NYA =5,133), had lower degrees (kOA =1.39 vs. kYA =1.66),
had lower clustering coefficients (COA =0.053 vs. CYA =0.115), and had larger shortest path
lengths (LOA =8.87 vs. LYA =7.45) than those of younger adults (see Table 2). Except for
one older adult, whose network had structural characteristics more similar to those of the
younger adults (row 5 in Table 2), these patterns held for all possible pairwise compar-
isons of younger and older adults’ networks. Also consistent with earlier work, the degree
distributions3of older adults’ networks were less similar to each other (.48 <r<.58) than
102 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
those of younger adults (.58 <r<.71), suggesting a progression of network differentiation
with age (Wulff et al., 2018).
5. Linking individual networks to cognitive performance
Can the networks described above be used to understand individual cognitive performance?
This question can be addressed at two levels: At the network level, macroscopic properties
of the network, such as average degree or clustering, could be used to predict overall per-
formance in a given task (Kenett & Faust, 2019). At the node level, microscopic properties,
such as node degree or the shortest path length between two nodes, can be used to predict
trial-level performance, such as the order of retrievals in a verbal fluency task. As the small
sample size (N=8) of our proof-of-concept study limits comparisons at the network level,
we focus on relating semantic networks and cognitive performance on the node level. To this
end, we analyze how well two important node characteristics—namely, node centrality, mea-
sured using PageRank (see, e.g., Griffiths et al., 2007), and node similarity, measured using
Katz’ walk similarity (see, e.g., De Deyne, Navarro, Perfors, & Storms, 2016; Hills, Jones, &
Todd, 2012)—correspond to response patterns in two verbal fluency tasks, a paired associative
learning task and an episodic memory task.
In the animal and letter verbal fluency tasks, individuals were asked to retrieve as many
animal words or words starting with the letter S, respectively, as they could, within 10 min.4
The retrieved words had substantially higher centrality than words that were not retrieved, but
are in an individual’s network. Specifically, for seven out of eight individuals, retrieved ani-
mals were more central in their respective networks as compared to other animals contained
in the respective networks that were not retrieved by the individual (Fig. 3A). Similarly, for
all eight individuals, retrievals of words starting with the letter S had higher centrality in the
respective networks than other words starting with any same letter contained in their net-
works (Fig. 3B). We also found words occurring directly adjacent within retrieval sequences
to be more similar to each other than words more distant from each other. Comparing directly
adjacent words (Lag 1) to words that were three responses apart, this was the case for every
individual in the animal fluency task (Fig. 3C) and for six of the eight individuals in the letter
fluency task (Fig. 3D). In the episodic memory task, participants freely recalled words from
previously studied word lists.5Both retrieved words and intrusions, that is retrieved words
that were not on the list, but included in the individual’s network, tended to have, on average,
higher centrality than missing words (Fig. 3E). This pattern held for five out of eight individu-
als. Furthermore, both retrieved words and intrusions were, on average, more similar to other
retrieved words than to missing words (Fig. 3F) for seven of eight individuals. Finally, the
associate recall task required individuals to retrieve previously learned word pairs.6Retrieved
word pairs had, on average and for all eight individuals, higher similarity than words pairs
that were not retrieved (Fig. 3G).
Each of these links demonstrates that individual networks can be used to predict an indi-
vidual’s pattern of behavior in several cognitive measures. The critical question arising from
this framework, however, is whether individual networks reveal idiosyncratic differences that
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 103
(A) (B) (C) (D)
(E) (F) (G)
Fig. 3. Links between mental representations and cognitive performance for the four younger (blue lines) and
the four older adults (yellow lines). The panels show the difference in PageRank centrality between retrieved
and nonretrieved words in animal (A) and letter fluency (B), the average cosine similarity of words one,
two, and three words apart in animal (C) and letter fluency (D), the PageRank centrality of retrieved, missing,
and intrusion words in the episodic memory task (E), the average cosine similarity of retrieved, missing, and
intrusion words to (other) retrieved words in the episodic memory task (F), and the average cosine similarity of
retrieved versus nonretrieved word pairs in the associative recall task (G). The black line shows the effects under
the aggregate representations. Centrality is measured as PageRank times the number of nodes in the network, to
account for differences in network sizes.
can be used to understand an individual’s behavior better than would be possible using an
aggregate representation derived from the responses of many individuals. To assess this,
we compared the magnitude of the effects presented in Fig. 3, for example, the PageRank
of retrieved versus nonretrieved words in verbal fluency tasks, which were derived using
individuals’ personal networks, against the magnitude of effects resulting from using an
aggregate derived from combining the responses from all eight individuals. This analysis
revealed that the aggregate network produced, on average, equally large, if not larger, effects
than individual networks (see black lines in Fig. 3).
The similar or improved performance of the aggregate compared to individuals’ own net-
works could be due to differences in both the reliability and precision of measurement of the
different types of networks. First, on average, only 53% of three possible associations given
to repeated cues were also given during the first encounter of the same cue, indicating, at best,
moderate levels of internal consistency. Furthermore, focusing on the 2,111 shared nodes, we
found individuals’ networks to share a large amount of variance with the aggregate network,
as indicated by average correlations of ¯r=.61 and ¯r=.55 for PageRank and cosine distri-
butions, respectively, although variance shared with the networks of other individuals was,
on average, rather low (PageRank: ¯r=.34; Cosine: ¯r=.33). These results suggest that the
104 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
individual networks contain a shared component that is reliably captured by an aggregate
network. They also suggest that the aggregate network is relatively free from aggregation
bias (see also Wulff et al., 2018). Second, the aggregate network derived from the response
of all individuals is given access to more detailed information on the relative strength of
associations between nodes in the networks. For individuals, only 2.2% to 5.5% of word
associations occurred more than once, and these repetitions occurred a maximum of two or
three times per individual. For the aggregate representation, however, 20.5% of associations
occurred more than once, with counts reaching as high as 15, leading to a more graded pattern
of edge strengths between network nodes in the aggregate relative to the individual networks.
One should note that this implies that the comparison of individual and aggregate networks
is largely confounded with a comparison of weighted (aggregate) and unweighted (individ-
ual) networks.
6. What have we learned from MySWOW so far?
Our approach using network science was successful in mapping large-scale individual net-
works for a small sample of younger and older adults, suggesting that the approach can be
used to capture the mental representations of individuals: Individuals’ networks were highly
similar to each other in terms of content, and these contents were qualitatively similar to
those of past efforts to produce aggregate networks, such as SWOW. In line with past work,
the results from the eight participants suggested age-related differences in their structural
composition (Dubossarsky et al., 2017; Wulff et al., 2018). All in all, our approach seems
promising in offering a window into adult age differences in the content and structure of the
aging lexicon and the individual differences that can arise across adulthood as a function of
cumulative experience.
Our results also showed that estimates of node centrality and relatedness derived from
single individuals’ networks exhibited strong links to individuals’ responses in four cog-
nitive tasks, including two verbal fluency and two memory tasks. These results demon-
strate that it is feasible, in principle, to map information from individual semantic networks
onto additional linguistic and memory performance at the single participant level. Naturally,
future work needs to strive to collect data from more individuals to assess whether network-
level properties are systematically related to differences in cognitive performance between
Our results demonstrate that our approach is, nevertheless, limited. Most crucially, our
results suggest that individual networks were less powerful in accounting for individuals’
response patterns in the additional measures of linguistic and memory performance than an
aggregate network composed of all individuals’ responses. The advantage of aggregate over
individual networks presents a major challenge to the claim that it is important to assess
individual networks to understand individual differences in cognitive performance. Before
one fully dismisses our claim, one should distinguish two explanations for these findings
that have different implications. First, one possibility is that our approach was not able to
capture relevant individual differences in the content of the semantic representations, for
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 105
example, because of the use of a pool of cues containing mostly high-frequency words for
which individual differences in semantic representations may be small. In this case, the
breadth of cues used would need to be expanded so that the networks generated are better
able to capture the idiosyncratic nature of individuals’ representations. Second, another pos-
sibility is that our approach may not have been able to measure the strength of association
between various nodes (i.e., edges) with sufficient precision and reliability. In this case, addi-
tional work may be needed that considers both obtaining larger sets of responses per cue,
having participants provide responses to the same cue on repeated occasions, or asking indi-
viduals to provide additional ratings (Roads & Love, 2020). In our view, these possibilities
speak to the need for future work that guarantees appropriate scope and reliability rather than
a clear rejection of the idea that individual networks can be helpful to understanding the link
between experience and cognitive performance.
7. General discussion
We proposed an empirical agenda that utilizes network science to provide an estimate of
the extent to which idiosyncratic experiences matter for the size and structure of mental rep-
resentations and, ultimately, individual and age differences in cognitive performance. Despite
the promise we see in this approach, and some encouraging results from our feasibility study,
a number of major challenges remain.
First, there are still considerable questions about how to best capture individual seman-
tic networks. As discussed in Section 1, the MySWOW approach is only one of several:
There are alternative elicitation methods (see Table 1) as well as different network estima-
tion methods (e.g., Zemla & Austerweil, 2018) that merit further consideration. Future work
may want to rely on several of the available methods and compare commonalities and dif-
ferences of their results to obtain a better picture of individual networks. One limitation that
may be common to many of the currently available approaches, however, is that no elicitation
of individuals’ knowledge store is independent of the process by which the representation
is accessed. As a result, any individual or age differences detected cannot be unequivocally
assigned to the nature of association in the knowledge store (representation), but can in prin-
ciple result from the process by which this representation is searched and accessed (Jones,
Hills, & Todd, 2015; Kenett, Beckage, Siew, & Wulff, 2020; Kraemer, Wulff, & Gluth, 2021;
Siew, Wulff, Beckage, & Kenett, 2019). So far, there seems to be no clear consensus concern-
ing the extent to which representation and process are entangled in the kinds of tasks typically
used to elicit semantic networks (Abbott, Austerweil, & Griffiths, 2015; Jones et al., 2015).
The degree of entanglement likely depends on the nature of the tasks and processes com-
pared. At least in word association tasks, empirical evidence suggests associative strength
between word pairs in itself is a poor predictor of a wide variety of findings that involve
retrieval about these pairs from semantic memory (De Deyne et al., 2019). The reason for this
is that the same retrieval processes underestimate semantic effects due to frequency biases
(i.e., a frequent word like money is the response to many cue words, and therefore car-
ries little information). In most studies, such biases need to be mitigated using appropriate
106 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
transformations of associative strength that highlight cue-specific information (e.g., point-
wise mutual information; De Deyne et al., 2019). More generally, it is unlikely that a single
approach can provide a definitive resolution to this conundrum and there are two possible
ways forward that we would like to emphasize. One potential way to distinguish the role of
structure versus process could be to use multiple methods of assessment (e.g., word associ-
ation, verbal fluency, relatedness judgments) to elicit mental representations. To the extent
that these different assessment methods are associated with different search processes and
strategies, convergent evidence could provide some support for the role of representation. A
related approach could involve the use of the same elicitation method under different explic-
itly instructed search strategies that could equate these across participants (Wulff, Hills, &
Hertwig, 2013, 2020). A second direction could involve using techniques from neuroscience
to distinguish between search (i.e., control) processes and representational components (Hoff-
man & Morcom, 2018), in particular to the extent that individuals’ representational space has
a signature in the functional organization of the brain (Huth, Nishimoto, Vu, & Gallant, 2012).
Second, we are still lacking a computational model of learning and cognition that captures
and links all aspects of interest (i.e., environment–representation–performance). A suite of
models would be ideal because they could be used to generate predictions about which tests
should be used to assess individual and age differences in mental representations as a function
of experience. These predictions could then guide specific empirical approaches and tests.
These models may also be generative in allowing us to test specific components contributing
to individual differences, such as the role of learning, memory, or search strategies (Wulff
et al., 2019). For example, past work has used techniques from network science that assess the
robustness of networks to perturbation and decay that could be instrumental in understanding
the role of aging processes, such as forgetting (Borge-Holthoefer, Moreno, & Arenas, 2011;
Cosgrove et al., 2021).
Third, there is still a dearth of data concerning single individuals’ exposure to the phys-
ical, linguistic, and social environment that can then be linked to individuals’ mental rep-
resentations. Most past work has used simple measures of self-reported exposure (e.g., to
print; Mol & Bus, 2011) but there may be more efficient alternatives that are now avail-
able, for example, experience sampling (Dennis, Yim, Garrett, Sreekumar, & Stone, 2019),
or massive recording and analysis of naturalistic linguistic exposure, such as speech (Mehl,
2017) and text (Banda et al., 2021). These approaches could become particularly powerful as
such data become increasingly connected with other information about individuals, for exam-
ple, if these are collected with informed consent in larger surveys or longitudinal household
panels that also include dedicated cognitive instruments, such as verbal fluency or memory
tasks (Taler, Johns, & Jones, 2020).
Fourth, and finally, most of past research and our own approach introduced above is cor-
relational in nature in that it aims to establish a correlation between individual experience
and the contents and structure of mental representation or cognitive performance. In order to
help establish causality, however, one would need a comparison of individuals assigned to
different environmental exposures through either natural or, ideally, controlled experiments.
There are a few candidate training strategies, including training of specific physical or virtual
environments (Miller et al., 2013), artificial or natural languages (Pothos, 2007), and even
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 107
complex narratives (Heusser, Fitzpatrick, & Manning, 2021). Regardless of the exact type of
information and mode of exposure, it remains a challenge to obtain meaningful and reliable
individual estimates of the emergent mental representations and their potential effects across
large swathes of time such as a life span.
8. Conclusion
We have argued that quantifying individual and age differences in the size and structure of
human knowledge is important because this represents a missing link in estimating the role
of cumulative experience in cognitive performance. We specifically proposed an empirical
agenda that combines tracking individuals’ idiosyncratic experiences, broad mapping of their
mental representations using the tools of network science, and studies linking representational
structure to individuals’ cognitive performance. We strongly believe such steps can move us
closer to understanding the role of the environment in shaping the structure of the mental
lexicon and its implications for cognitive aging.
We are grateful to Laura Wiles for editing the article. This work was supported by a
grant from the Swiss National Science Foundation ( to Dirk
U. Wulff.
Open access funding provided by Universitat Basel.
Open Research Badges
This article has earned Open Data. Data is available at
1 A recent alternative approach (Roads & Love, 2020) that relies on ranking and active
learning algorithms can provide a solution to the problem of scale.
2 German SWOW data were downloaded on January 25, 2021.
3 Considering only the 2,111 words shared between all individual networks.
4 On average, between 62 and 113 animals and between 45 and 138 words starting with the
letter S were retrieved. Animals overlapped with 1.5% of cues and 0.8% of responses,
whereas words starting with the letter S overlapped with 11.1% of cues and 11.9% of
responses. For details see Wulff et al. (in press).
5 Participants correctly recalled between 28.7% and 60.9% of words, with an additional
1.3% to 25% intrusions. For details see Wulff et al. (in press).
6 Participants correctly recalled between 32.8% and 96.8% of pairs. For details see Wulff
et al. (in press).
108 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
Abbott, J. T., Austerweil, J. L., & Griffiths, T. L. (2015). Random walks on semantic networks can resemble
optimal foraging. Psychological Review,122(3), 558–569.
Banda, J. M., Tekumalla, R., Wang, G., Yu, J., Liu, T., Ding, Y., Artemova, E., Tutubalina, E., & Chowell, G.
(2021). A large-scale COVID-19 twitter chatter dataset for open scientific research—An international collabo-
ration. Epidemiologia,2(3), 315–324.
Benedek, M., Kenett, Y. N., Umdasch, K., Anaki, D., Faust, M., & Neubauer, A. C. (2017). How semantic memory
structure and intelligence contribute to creative thought: A network science approach. Thinking & Reasoning,
23(2), 158–183.
Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of communities in large
networks. Journal of Statistical Mechanics: Theory and Experiment,2008(10), Article P10008.
Borge-Holthoefer, J., Moreno, Y., & Arenas, A. (2011). Modeling abnormal priming in Alzheimer’s patients with
a free association network. PLOS ONE,6(8), Article e22651.
Brysbaert, M., Stevens, M., Mandera, P., & Keuleers, E. (2016). How many words do we know? Practical esti-
mates of vocabulary size dependent on word definition, the degree of language input and the participant’s age.
Frontiers in Psychology,7, Article 1116.
Buchler, N. E. G., & Reder, L. M. (2007). Modeling age-related memory deficits: A two-parameter solution.
Psychology and Aging,22(1), 104–121.
Cosgrove, A. L., Kenett, Y. N., Beaty, R. E., & Diaz, M. T. (2021). Quantifying flexibility in thought: The resiliency
of semantic networks differs across the lifespan. Cognition,211, Article 104631.
De Deyne, S., Navarro, D. J., Perfors, A., & Storms, G. (2016). Structure at every scale: A semantic network
account of the similarities between unrelated concepts. Journal of Experimental Psychology: General,145(9),
De Deyne, S., Navarro, D. J., & Storms, G. (2013). Better explanations of lexical and semantic cognition using
networks derived from continued rather than single-word associations. Behavior Research Methods,45(2),
De Deyne, S., Navarro, D. J., Collell, G., & Perfors, A. (2021). Visual and affective multimodal models of word
meaning in language and mind. Cognitive Science,45(1), Article e12922.
De Deyne, S., Navarro, D. J., Perfors, A., Brysbaert, M., & Storms, G. (2019). The “small world of words”
English word association norms for over 12,000 cue words. Behavior Research Methods,51(3), 987–1006.
De Deyne, S., & Storms, G. (2008). Word associations: Norms for 1,424 Dutch words in a continuous task.
Behavior Research Methods,40(1), 198–205.
Dennis, S., Yim, H., Garrett, P., Sreekumar, V., & Stone, B. (2019). A system for collecting and analyzing
experience-sampling data. Behavior Research Methods,51(4), 1824–1838.
Dubossarsky, H., De Deyne, S., & Hills, T. T. (2017). Quantifying the structure of free association networks across
the life span. Developmental Psychology,53(8), 1560–1570.
Gobet, F., & Simon, H. A. (1996). Recall of random and distorted chess positions: Implications for the theory of
expertise. Memory & Cognition,24(4), 493–503.
Griffiths, T. L., Steyvers, M., & Firl, A. (2007). Google and the mind: Predicting fluency with pagerank. Psycho-
logical Science,18(12), 1069–1076.
He, L., Kenett, Y. N., Zhuang, K., Liu, C., Zeng, R., Yan, T., Huo, T., & Qiu, J. (2021). The relation between
semantic memory structure, associative abilities, and verbal and figural creativity. Thinking & Reasoning,27(2),
D. U. Wulff et al. /Topics in Cognitive Science 14 (2022) 109
Heusser, A. C., Fitzpatrick, P. C., & Manning, J. R. (2021). Geometric models reveal behavioural and neural
signatures of transforming experiences into memories. Nature Human Behaviour,5(7), 905–919. https://doi.
Hills, T. T., Jones, M. N., & Todd, P. M. (2012). Optimal foraging in semantic memory. Psychological Review,
119(2), 431–440.
Hoffman, P., & Morcom, A. M. (2018). Age-related changes in the neural networks supporting semantic cognition:
A meta-analysis of 47 functional neuroimaging studies. Neuroscience & Biobehavioral Reviews,84, 134–150.
Huettig, F., & Pickering, M. J. (2019). Literacy advantages beyond reading: Prediction of spoken language. Trends
in Cognitive Sciences,23(6), 464–475.
Huth, A. G., Nishimoto, S., Vu, A. T., & Gallant, J. L. (2012). A continuous semantic space describes the rep-
resentation of thousands of object and action categories across the human brain. Neuron,76(6), 1210–1224.
Jones, M. N., Hills, T. T., & Todd, P. M. (2015). Hidden processes in structural representations: A reply to Abbott,
Austerweil, and Griffiths (2015). Psychological Review,122(3), 570–574.
Kenett, Y. N., Beckage, N. M., Siew, C. S. Q., & Wulff, D. U. (2020). Cognitive network science: A new frontier.
Complexity,2020, Article e6870278.
Kenett, Y. N., & Faust, M. (2019). A semantic network cartography of the creative mind. Trends in Cognitive
Sciences,23(4), 271–274.
Kraemer, P., Wulff, D. U., & Gluth, S. (2021). A sequential sampling account of semantic relatedness decisions.
Lindenberger, U. (2014). Human cognitive aging: Corriger la fortune? [Publisher: American Association for the
Advancement of Science]. Science,346(6209), 572–578.
Love, R., Dembry, C., Hardie, A., Brezina, V., & McEnery, T. (2017). The spoken BNC2014: Designing and
building a spoken corpus of everyday conversations. International Journal of Corpus Linguistics,22(3), 319–
Mehl, M. R. (2017). The electronically activated recorder (EAR): A method for the naturalistic observation of
daily social behavior. Current Directions in Psychological Science,26(2), 184–190.
Miller, J. F., Neufang, M., Solway, A., Brandt, A., Trippel, M., Mader, I., Hefft, S., Merkow, M., Polyn, S. M.,
Jacobs, J., Kahana, M. J., & Schulze-Bonhage, A. (2013). Neural activity in human hippocampal formation
reveals the spatial context of retrieved memories. Science,342(6162), 1111–1114.
Mol, S. E., & Bus, A. G. (2011). To read or not to read: A meta-analysis of print exposure from infancy to early
adulthood. Psychological Bulletin,137(2), 267–296.
Morais, A. S., Olsson, H., & Schooler, L. J. (2013). Mapping the structure of semantic memory. Cognitive Science,
37(1), 125–145.
Nation, K. (2017). Nurturing a lexical legacy: Reading experience is critical for the development of word reading
skill. NPJ Science of Learning,2(1), 1–4.
Nelson, D. L., Bennett, D. J., Gee, N. R., Schreiber, T. A., & McKinney, V. M. (1993). Implicit memory: Effects
of network size and interconnectivity on cued recall. Journal of Experimental Psychology: Learning, Memory,
and Cognition,19(4), 747–764.
Pothos, E. M. (2007). Theories of artificial grammar learning. Psychological Bulletin,133(2), 227–244. https:
Ramscar, M., Hendrix, P., Shaoul, C., Milin, P., & Baayen, H. (2014). The myth of cognitive decline: Non-linear
dynamics of lifelong learning. Topics in Cognitive Science,6(1), 5–42.
Ramscar, M., Sun, C. C., Hendrix, P., & Baayen, H. (2017). The mismeasurement of mind: Life-span changes
in paired-associate-learning scores reflect the “cost” of learning, not cognitive decline. Psychological Science,
28(8), 1171–1179.
110 D. U. Wulff et al. /Topics in Cognitive Science 14 (2022)
Roads, B. D., & Love, B. C. (2020). Enriching imagenet with human similarity judgments and psychological
embeddings. arXiv.
Rodd, J. M. (2020). Settling into semantic space: An ambiguity-focused account of word-meaning access. Per -
spectives on Psychological Science,15(2), 411–427.
Sala, G., & Gobet, F. (2017). Experts’ memory superiority for domain-specific random material generalizes across
fields of expertise: A meta-analysis. Memory & Cognition,45(2), 183–193.
Schröter, P., & Schroeder, S. (2017). The developmental lexicon project: A behavioral database to investigate
visual word recognition across the lifespan. Behavior Research Methods,49(6), 2183–2203.
Siew, C. S. Q., Wulff, D. U., Beckage, N. M., & Kenett, Y. N. (2019). Cognitive network science: A review of
research on cognition through the lens of network representations, processes, and dynamics. Complexity,2019,
Article e2108423.
Steyvers, M., Shiffrin, R. M., & Nelson, D. L. (2005). Word association spaces for predicting semantic similarity
effects in episodic memory. In Alice F Healey (ed.), Experimental cognitive psychology and its applications,
Decade of behavior (pp. 237–249). Washington, D.C., U.S. American Psychological Association. https://doi.
Taler, V., Johns, B. T., & Jones, M. N. (2020). A large-scale semantic analysis of verbal fluency across the aging
spectrum: Data from the Canadian longitudinal study on aging. The Journals of Gerontology: Series B,75(9),
Twenge, J. M., Martin, G. N., & Spitzberg, B. H. (2019). Trends in U.S. Adolescents’ media use, 1976–2016: The
rise of digital media, the decline of TV, and the (near) demise of print. Psychology of Popular Media Culture,
8(4), 329–345.
Verhaeghen, P. (2003). Aging and vocabulary score: A meta-analysis. Psychology and Aging,18(2), 332–339.
Wlotko, E. W., Federmeier, K. D., & Kutas, M. (2012). To predict or not to predict: Age-related differences in the
use of sentential context. Psychology and Aging,27(4), 975–988.
Wulff, D. U., Aeschbach, S., De Deyne, S., & Mata, R. (in press). Data from the MySWOW proof-of-concept
study: Linking individualsemantic networks and cognitive performance. Journal of Open Psychology Data.
Wulff, D. U., De Deyne, S., Jones, M. N., Mata, R., & The Aging Lexicon Consortium. (2019). New perspectives
on the aging lexicon. Trends in Cognitive Sciences,23(8), 686–698.
Wulff, D. U., Hills, T., & Hertwig, R. (2020). Memory is one representation not many: Evidence against wormholes
in memory. PsyArXiv.
Wulff, D. U., Hills, T., & Mata, R. (2018). Structural differences in the semantic networks of younger and older
adults. PsyArXiv.
Wulff, D. U., Hills, T. T., & Hertwig, R. (2013). Worm holes in memory: Is memory one representation or
many? Proceedings of the Annual Meeting of the Cognitive Science Society,35.
Wulff, D. U., Hills, T. T., Lachman, M., & Mata, R. (2016). The aging lexicon: Differences in the semantic
networks of younger and older adults. Proceedings of the Annual Meeting of the Cognitive Science Society,38.
Zemla, J. C., & Austerweil, J. L. (2018). Estimating semantic networks of groups and individuals from fluency
data. Computational Brain & Behavior,1(1), 36–58. 7
... Semantic memory network analyses have been utilized across various research areas-creativity, language acquisition, clinical populations, and in healthy aging (Castro et al., 2020;Kenett & Faust, 2019b;Steyvers & Tenenbaum, 2005;Wulff et al., 2019). Previous work that focused on semantic memory networks and aging suggests that with increased age, semantic memory becomes less efficient, organized, and connected (Cosgrove et al., 2021;Dubossarsky et al., 2017;Kenett et al., 2021;Wulff et al., 2019;Wulff, De Deyne, Aeschbach, et al., 2022;Wulff, Hills, & Mata, 2022). ...
... Although many studies have examined group differences in network structure and others have emphasized the importance of individual differences in network structure relevant for memory retrieval and creative thinking (Kenett & Faust, 2019b;Siew et al., 2019;Zemla & Austerweil, 2018), few studies have looked at individual network differences in aging Wulff, De Deyne, Aeschbach, et al., 2022;Wulff, Hills, & Mata, 2022). In one such study, Wulff et al. calculated individual semantic networks using semantic similarity ratings (Wulff et al., 2018). ...
... While they did not relate these structural differences to cognitive behavior, the greater variance among older adults in the similarity ratings was likely related to differences in life experiences and acquired knowledge Wulff, Hills, & Mata, 2022). Although the researchers conclude that the individual networks were less powerful in capturing the linguistic response patterns than the aggregated semantic memory network, this study had a relatively small sample size (n = 8, Wulff, De Deyne, Aeschbach, et al., 2022). Since there were only four younger adults and four older adults, the individual-based variation in network structure that accounts for behavioral performance on language-related tasks may not have been sufficient to reveal significant network effects. ...
Computational research suggests that semantic memory, operationalized as semantic memory networks, undergoes age-related changes. Previous work suggests that concepts in older adults' semantic memory networks are more separated, more segregated, and less connected to each other. However, cognitive network research often relies on group averages (e.g., young vs. older adults), and it remains unclear if individual differences influence age-related disparities in language production abilities. Here, we analyze the properties of younger and older participants' individual-based semantic memory networks based on their semantic relatedness judgments. We related individual-based network measures-clustering coefficient (CC; connectivity), global efficiency, and modularity (structure)-to language production (verbal fluency) and vocabulary knowledge. Similar to previous findings, we found significant age effects: CC and global efficiency were lower, and modularity was higher, for older adults. Furthermore, vocabulary knowledge was significantly related to the semantic memory network measures: corresponding with the age effects, CC and global efficiency had a negative relationship, while modularity had a positive relationship with vocabulary knowledge. More generally, vocabulary knowledge significantly increased with age, which may reflect the critical role that the accumulation of knowledge within semantic memory has on its structure. These results highlight the impact of diverse life experiences on older adults' semantic memory and demonstrate the importance of accounting for individual differences in the aging mental lexicon. (PsycInfo Database Record (c) 2023 APA, all rights reserved).
... Past work has often made the simplifying assumption that a common semantic network can be used to understand human semantic cognition 4,5,[8][9][10][11] . This assumptions is implicit, for instance, in efforts to model retrieval from memory 12 , judgments of relatedness 13 , or decision making 14 using large-scale word vector spaces and free-association networks. However, general theories of learning and development 15,16 , as well as empirical findings [17][18][19] , suggest that semantic networks could vary considerably between individuals and across the life span. ...
... More recently, however, research suggests that individual learning and life span development can also lead to changes in the structure of human knowledge [23][24][25] . For example, recent efforts have used data from large-scale free-association studies to show that older adults' semantic networks are less connected, efficient, and structured relative to those of younger adults 12,18 . ...
... Despite such differences, the age-related patterns in macroscopic network structure generalize across the different domains and conditions, which speaks to the generality of these findings across elicitation procedures (cf. 12,18 ). ...
Full-text available
Cognitive science invokes semantic networks to explain diverse phenomena, from memory retrieval to creativity. Research in these areas often assumes a single underlying semantic network that is shared across individuals. Yet, recent evidence suggests that content, size, and connectivity of semantic networks are experience-dependent, implying sizable individual and age-related differences. Here, we investigate individual and age differences in the semantic networks of younger and older adults by deriving semantic networks from both fluency and similarity rating tasks. Crucially, we use a megastudy approach to obtain thousands of similarity ratings per individual to allow us to capture the characteristics of individual semantic networks. We find that older adults possess lexical networks with smaller average degree and longer path lengths relative to those of younger adults, with older adults showing less interindividual agreement and thus more unique lexical representations relative to younger adults. Furthermore, this approach shows that individual and age differences are not evenly distributed but, rather, are related to weakly connected, peripheral parts of the networks. All in all, these results reveal the interindividual differences in both the content and the structure of semantic networks that may accumulate across the life span as a function of idiosyncratic experiences.
... One database is the Small World of Words [34] projects, collecting word associations online. However, it is unknown exactly how many brand names are part of the database; a random search of the available online data suggests that at least some car brand names are part of the database (for some results, see [35,36]). Another somewhat older data collection is from Human Brain Cloud [37], which has collected data since 2007. ...
... All in all, word association networks can be used for the description of several cognitive tasks, for example, predicting lexical norms [39], memory recall tasks [40], change of associative networks across the life span [35,41] or description of the mental lexicon's multilayer characteristics [36]. economics and linguistics [42]. ...
Full-text available
Brands can be defined as psychological constructs residing in our minds. By analyzing brand associations, we can study the mental constructs around them. In this paper, we study brands as parts of an associative network based on a word association database. We explore the communities–closely-knit groups in the mind–around brand names in this structure using two community detection algorithms in the Hungarian word association database ConnectYourMind. We identify brand names inside the communities of a word association network and explain why these brand names are part of the community. Several detected communities contain brand names from the same product category, and the words in these categories were connected either to brands in the category or to words describing the product category. Based on our findings, we describe the mental position of brand names. We show that brand knowledge, product knowledge and real word knowledge interact with each other. We also show how the meaning of a product category arises and how this meaning is related to brand meaning. Our results suggest that words sharing the same community with brand names can be used in brand communication and brand positioning.
... Second, our elicitation method cannot provide truly individual semantic representations. For this purpose, richer sets of data would be needed, requiring more intensive designs [e.g., (42)(43)(44)]. Such approaches would be instrumental to improving predictions at the individual level, thus fulfilling the promise of uncovering the role of semantic representations for individual differences in risk-related constructs and other domains. ...
... Recent work has suggested that lexical network approaches may be powerful tools to make predictions about population-level risk perception of novel risks (45), and it would be interesting to assess the possibility of extending our approach to make similar but personalized predictions for single individuals across all areas of knowledge [cf. (43)]. ...
Full-text available
What are the defining features of lay people’s semantic representation of risk? We contribute to mapping the semantics of risk based on word associations to provide insight into both universal and individual differences in the representation of risk. Specifically, we introduce a mini-snowball word association paradigm and use the tools of network and sentiment analysis to characterize the semantics of risk. We find that association-based representations not only corroborate but also extend those extracted from past survey- and text-based approaches. Crucially, we find that the semantics of risk show universal properties and individual and group differences. Most notably, while semantic clusters generalize across languages, their frequency varies systematically across demographic groups, with older and female respondents showing more negative connotations and mentioning more often certain types of activities (e.g., recreational activities) relative to younger adults and males, respectively. Our work has general implications for the measurement of risk-related constructs by suggesting that “risk” can mean different things to different individuals.
... The first step in any network analysis is constructing the network that represents the studied domain. Using networks based on large language corpora as inspiration (e.g., De Deyne, Navarro, Perfors, Brysbaert, & Storms, 2019; Levy et al., 2021;Stella & Brede, 2016;Wulff, De Deyne, Aeschbach, & Mata, 2022), we construct a melodic network based on a large corpus of existing music. Musical improvisation is produced under real-time constraints. ...
Music is a complex system consisting of many dimensions and hierarchically organized information-the organization of which, to date, we do not fully understand. Network science provides a powerful approach to representing such complex systems, from the social networks of people to modelling the underlying network structures of different cognitive mechanisms. In the present research, we explored whether network science methodology can be extended to model the melodic patterns underlying expert improvised music. Using a large corpus of transcribed improvisations, we constructed a network model in which 5-pitch sequences were linked depending on consecutive occurrences, constituting 116,403 nodes (sequences) and 157,429 edges connecting them. We then investigated whether mathematical graph modelling relates to musical characteristics in real-world listening situations via a behavioral experiment paralleling those used to examine language. We found that as melodic distance within the network increased, participants judged melodic sequences as less related. Moreover, the relationship between distance and reaction time (RT) judgements was quadratic: participants slowed in RT up to distance four, then accelerated; a parallel finding to research in language networks. This study offers insights into the hidden network structure of improvised tonal music and suggests that humans are sensitive to the property of melodic distance in this network. More generally, our work demonstrates the similarity between music and language as complex systems, and how network science methods can be used to quantify different aspects of its complexity.
... Understanding the structure and organisation of knowledge in the mental lexicon requires a framework that is quantitative [27], interpretable [28] and human-centric [29]. This framework must: (i) be capable of producing inferences and comparable measurements regulated by mathematical equations and theoretical models [16,30] (quantitative); (ii) map results to outputs through an internal representation of knowledge available to researchers, unlike most black-box machine learning knowledge models [21] (interpretable); and (iii) be grounded in psychological theory and large-scale datasets in order to account for the complex nuances of human psychology rather than make abstract inferences that are of little value to psychologists [31,32]. An artificial intelligence that categorises individuals using binary labels like "aphasic" or "healthy" without identifying the severity of their language impairments, nor considers their ability to acquire, retain, and produce new knowledge would not be human-centric [17]). ...
Full-text available
The mental lexicon is a complex cognitive system representing information about the words/concepts that one knows. Decades of psychological experiments have shown that conceptual associations across multiple, interactive cognitive levels can greatly influence word acquisition, storage, and processing. How can semantic, phonological, syntactic, and other types of conceptual associations be mapped within a coherent mathematical framework to study how the mental lexicon works? We here review cognitive multilayer networks as a promising quantitative and interpretative framework for investigating the mental lexicon. Cognitive multilayer networks can map multiple types of information at once, thus capturing how different layers of associations might co-exist within the mental lexicon and influence cognitive processing. This review starts with a gentle introduction to the structure and formalism of multilayer networks. We then discuss quantitative mechanisms of psychological phenomena that could not be observed in single-layer networks and were only unveiled by combining multiple layers of the lexicon: (i) multiplex viability highlights language kernels and facilitative effects of knowledge processing in healthy and clinical populations; (ii) multilayer community detection enables contextual meaning reconstruction depending on psycholinguistic features; (iii) layer analysis can mediate latent interactions of mediation, suppression and facilitation for lexical access. By outlining novel quantitative perspectives where multilayer networks can shed light on cognitive knowledge representations, also in next-generation brain/mind models, we discuss key limitations and promising directions for cutting-edge future research.
... The first step in any network analysis is constructing the network that represents the studied domain. Using networks based on large language corpora as inspiration (e.g., De Deyne et al., 2019; Levy et al., 2021;Stella & Brede, 2016;Wulff et al., 2022), we construct a melodic network based on a large corpus of existing music. Musical improvisation is produced under real-time constraints. ...
Full-text available
Music is a complex system consisting of many dimensions and hierarchically organized information—the organization of which, to date, we do not fully understand. Network science provides a powerful approach to representing such complex systems, from the social networks of people to modelling the underlying network structures of different cognitive mechanisms. In the present research, we explored whether network science methodology can be extended to model the melodic patterns underlying expert improvised music. Using a large corpus of transcribed improvisations, we constructed a network model in which 5-pitch sequences were linked depending on consecutive occurrences, constituting 116,407 nodes (sequences) and 157,429 edges connecting them. We found that the network exhibited structural properties that resemble “scale-free” networks (i.e., networks with degree distribution following a power law). We then investigated whether mathematical graph modeling relates to musical characteristics in real-world listening situations via a behavioral experiment paralleling those used to construct semantic networks in language. We found that as distance within the network increased, participants judged melodic sequences as less related. Moreover, the relationship between distance and reaction time (RT) judgments was quadratic: participants slowed in RT up to distance four, then accelerated; a parallel finding to research in language networks. This study offers insights into the hidden network structure of improvised tonal music and suggests that humans are sensitive to the property of melodic distance in this network.
... While estimating group-based semantic memory networks from semantic fluency data is a common practice , we acknowledge the need for conducting future similar studies which will estimate individual-based semantic networks, in relation to creativity and bilingualism. Such future studies can utilize methods that are currentlyy being developed to estimate individual-based semantic memory networks (He et al., 2021;Ovando-Tellez et al., 2022;Wulff et al., 2022). ...
Full-text available
Creativity is related to a higher flexible semantic memory structure, which could explain greater fluency of ideas. Extensive research has identified a positive connection between creativity and bi-/multilingualism mainly in contexts where two languages or more concur in daily communicative interactions. Yet, creativity has received scant attention as regards L2 (second or foreign language) acquisition that mainly takes place in classroom situations. The scarce research points to a positive relationship between creativity and L2 fluency – understood as the number of words produced. We apply computational network science analysis and Forward Flow techniques to examine lexical organization patterns of a low creativity (LC) and high creativity (HC) group of 12th grade Spanish EFL learners. The participants completed two fluency tasks, where they generated animal names in their L2, and also L1 – used here as a control measure. EFL proficiency was controlled. Our analyses revealed that the HC individuals were more fluent in L1 and L2, generated more remote responses, and exhibited a more flexible and efficiently structured semantic memory in both languages, with a greater effect of creativity in L2. Contrary to previous research, the L2 semantic memory network exhibited a less random organization. Differences in the L2 learning conditions are adduced as likely causes of this result.
... Future studies are needed investigate the influence of individual differences in the language environment on the development of the semantic network of children with CI. As Wulff et al. (2022) have pointed out, research is needed to link how an individual's linguistic and social environment is shaping mental representation and thereby performance on cognitive (and language) tasks. Also, more studies are needed to investigate the development of multimodal semantic networks of children with CI using more than one language modality. ...
Purpose Kenett et al. (2013) report that the sematic networks, measured by using an oral semantic fluency task, of children with cochlear implants (CI) are less structured compared to the sematic networks of children with typical hearing (TH). This study aims to evaluate if such differences are only evident if children with CI are compared to children with TH matched on chronological age, or also if they are compared to children with TH matched on hearing age. Method The performance of a group of children with CI on a verbal fluency task was compared to the performance of a group of chronological-age matched children with TH. Subsequently, computational network analysis was used to compare the semantic network structure of the groups. The same procedure was applied to compare a group of children with CI to a group of hearing-age matched children with TH. Results The children with CI perform on the same level on an oral semantic verbal fluency task as the children with TH matched on hearing age. There are significant differences in terms of the structure of the semantic network between the groups. The magnitude of these differences is very small and they are non-significant for a proportion of nodes included in the bootstrap analysis. This indicates that there is no true difference between the networks. Hearing age, but not age at implantation was found to be significantly positively correlated with semantic verbal fluency performance for the children with CI. Conclusions The results from the current study indicate that length of language exposure is an important factor for the structure of the semantic network and the performance on a semantic verbal fluency task for children with CI. Further studies are needed to explore the role of the accessibility of the language input for the development of semantic networks of children with CI.
Conceptual knowledge is dynamic, fluid, and flexible, changing as a function of contextual factors at multiple scales. The Covid-19 pandemic can be considered a large-scale, global context that has fundamentally altered most people's experiences with the world. It has also introduced a new concept, COVID (or COVID-19), into our collective knowledgebase. What are the implications of this introduction for how existing conceptual knowledge is structured? Our collective emotional and social experiences with the world have been profoundly impacted by the Covid-19 pandemic, and experience-based perspectives on concept representation suggest that emotional and social experiences are critical components of conceptual knowledge. Such changes in collective experience should, then, have downstream consequences on knowledge of emotion- and social-related concepts. Using a naturally occurring dataset derived from the social media platform Twitter, we show that semantic spaces for concepts related to our emotional experiences with Covid-19 (i.e., emotional concepts like FEAR)-but not for unrelated concepts (i.e., animals like CAT)-show quantifiable shifts as a function of the emergence of COVID-19 as a concept and its associated emotional and social experiences, shifts which persist 6 months after the onset of the pandemic. The findings support a dynamic view of conceptual knowledge wherein shared experiences affect conceptual structure.
Full-text available
We report data from a proof-of-concept study involving the concurrent assessment of large-scale individual semantic networks and cognitive performance. The data include10,800 free associations—collected using a dedicated web-based platform over the course of several weeks—and responses to several cognitive tasks, including verbal fluency, episodic memory, associative recall tasks, from four younger and four older native German speakers. The data are unique in scope and composition and shed light on individual and age-related differences in mental representations and their role in cognitive performance across the lifespan.
Full-text available
We report data from a proof-of-concept study involving the concurrent assessment of large-scale individual semantic networks and cognitive performance. The data include 10,800 free associations-collected using a dedicated web-based platform over the course of 2-4 weeks-and responses to several cognitive tasks, including verbal fluency, episodic memory, associative recall tasks, from four younger and four older native German speakers. The data are unique in scope and composition and shed light on individual and age-related differences in mental representations and their role in cognitive performance across the lifespan.
Full-text available
Older adults tend to have a broader vocabulary compared to younger adults –indicating a richer storage of semantic knowledge – but their retrieval abilities decline with age. Recent advances in quantitative methods based on network science have investigated the effect of aging on semantic memory structure. However, it is yet to be determined how this aging effect on semantic memory structure relates to its overall flexibility. Percolation analysis provides a quantitative measure of the flexibility of a semantic network, by examining how a semantic memory network is resistant to “attacks” or breaking apart. In this study, we incorporated percolation analyses to examine how semantic networks of younger and older adults break apart to investigate potential age-related differences in language production. We applied the percolation analysis to 3 independent sets of data (total N = 78 younger, 78 older adults) from which we generated semantic networks based on verbal fluency performance. Across all 3 datasets, the percolation integrals of the younger adults were larger than older adults, indicating that older adults’ semantic networks were less flexible and broke down faster than the younger adults’. Our findings provide quantitative evidence for diminished flexibility in older adults’ semantic networks, despite the stability of semantic knowledge across the lifespan. This may be one contributing factor to age-related differences in language production.
Full-text available
How do we preserve and distort our ongoing experiences when encoding them into episodic memories? The mental contexts in which we interpret experiences are often person-specific, even when the experiences themselves are shared. Here we develop a geometric framework for mathematically characterizing the subjective conceptual content of dynamic naturalistic experiences. We model experiences and memories as trajectories through word-embedding spaces whose coordinates reflect the universe of thoughts under consideration. Memory encoding can then be modelled as geometrically preserving or distorting the ‘shape’ of the original experience. We applied our approach to data collected as participants watched and verbally recounted a television episode while undergoing functional neuroimaging. Participants’ recountings preserved coarse spatial properties (essential narrative elements) but not fine spatial scale (low-level) details of the episode’s trajectory. We also identified networks of brain structures sensitive to these trajectory shapes.
Full-text available
We compare three sequential sampling models, the Race model, the leaky competing accumulator model (LCA) and the drift diffusion model (DDM), as novel computational accounts of choices and response times in semantic relatedness decisions. We focus on two empirical benchmarks, the relatedness effect, denoting faster ”related” than ”unrelated” decisions when judging the relatedness of word pairs, and an inverted-U shaped relationship between response time and the relatedness strength of word pairs. Using simulations, we show that the LCA and DDM, but not the Race model, can reproduce both effects. Furthermore, we show that the LCA and DDM differ in their account of the relatedness effect, producing it under different circumstances and using different mechanisms. This observation offers a critical test, involving a novel, inverted relatedness effect for low-relatedness word pairs. Reanalyzing a publicly available data set, we obtained credible evidence of such an inverted relatedness effect, which is consistent with the LCA, but not the DDM. These results provide strong support for the LCA as an accurate computational account of semantic relatedness decisions and suggest an important role for decision-related processes in (semantic) memory tasks.
Full-text available
One of the main limitations of natural language‐based approaches to meaning is that they do not incorporate multimodal representations the way humans do. In this study, we evaluate how well different kinds of models account for people's representations of both concrete and abstract concepts. The models we compare include unimodal distributional linguistic models as well as multimodal models which combine linguistic with perceptual or affective information. There are two types of linguistic models: those based on text corpora and those derived from word association data. We present two new studies and a reanalysis of a series of previous studies. The studies demonstrate that both visual and affective multimodal models better capture behavior that reflects human representations than unimodal linguistic models. The size of the multimodal advantage depends on the nature of semantic representations involved, and it is especially pronounced for basic‐level concepts that belong to the same superordinate category. Additional visual and affective features improve the accuracy of linguistic models based on text corpora more than those based on word associations; this suggests systematic qualitative differences between what information is encoded in natural language versus what information is reflected in word associations. Altogether, our work presents new evidence that multimodal information is important for capturing both abstract and concrete words and that fully representing word meaning requires more than purely linguistic information. Implications for both embodied and distributional views of semantic representation are discussed.
Advances in supervised learning approaches to object recognition flourished in part because of the availability of high-quality datasets and associated benchmarks. However, these benchmarks—such as ILSVRC—are relatively task-specific, focusing predominately on predicting class labels. We introduce a publicly-available dataset that embodies the task-general capabilities of human perception and reasoning. The Human Similarity Judgments extension to ImageNet (ImageNet-HSJ) is composed of a large set of human similarity judgments that supplements the existing ILSVRC validation set. The new dataset supports a range of task and performance metrics, including evaluation of unsupervised algorithms. We demonstrate two methods of assessment: using the similarity judgments directly and using a psychological embedding trained on the similarity judgments. This embedding space contains an order of magnitude more points (i.e., images) than previous efforts based on human judgments. We were able to scale to the full 50,000 image ILSVRC validation set through a selective sampling process that used variational Bayesian inference and model ensembles to sample aspects of the embedding space that were most uncertain. To demonstrate the utility of ImageNet-HSJ, we used the similarity ratings and the embedding space to evaluate how well several popular models conform to human similarity judgments. One finding is that the more complex models that perform better on task-specific benchmarks do not better conform to human semantic judgments. In addition to the human similarity judgments, pre-trained psychological embeddings and code for inferring variational embeddings are made publicly available. ImageNet-HSJ supports the appraisal of internal representations and the development of more humanlike models.
Objectives The present study aimed to characterize changes in verbal fluency performance across the lifespan using data from the Canadian Longitudinal Study on Aging (CLSA). Methods We examined verbal fluency performance in a large sample of adults aged 45–85 (n = 12,686). Data are from the Tracking cohort of the CLSA. Participants completed a computer-assisted telephone interview that included an animal fluency task, in which they were asked to name as many animals as they could in 1 min. We employed a computational modeling approach to examine the factors driving performance on this task. Results We found that the sequence of items produced was best predicted by their semantic neighborhood, and that pairwise similarity accounted for most of the variance in participant analyses. Moreover, the total number of items produced declined slightly with age, and older participants produced items of higher frequency and denser semantic neighborhood than younger adults. Discussion These findings indicate subtle changes in the way people perform this task as they age. The use of computational models allowed for a large increase in the amount of variance accounted for in this data set over standard assessment types, providing important theoretical insights into the aging process.
Research has independently highlighted the roles of semantic memory and associative abilities in creative thinking. However, it remains unclear how these two capacities relate to each other, nor how they facilitate different creative thinking modalities, such as verbal and figural creativity. This study employed multiple cognitive tests and network science methodologies to shed light on the relationship between them. We constructed individual based semantic networks and assessed associative abilities, verbal and figural creative thinking. In line with previous studies, we found a relation between verbal creativity and more flexible semantic memory structure (higher connectivity, shorter distances between concepts, and lower modularity). However, we did not find any such relation between figural creativity and semantic memory structure. Associative abilities mediated the relationship between semantic memory structure and verbal creativity, implying the efficient spread of information in semantic memory may facilitate verbal creative thinking via associative abilities. These findings support and extend the associative theory of creativity and shed novel light on the relationship between semantic memory structure, associative abilities, and creativity.