Age differences in semantic network structure: acquiring knowledge shapes semantic memory

Abstract

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 (connectivity), global efficiency, and modularity (structure)—to language production (verbal fluency) and vocabulary knowledge. Similar to previous findings, we found significant age effects: clustering coefficient 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, clustering coefficient 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.

Introduction

Healthy aging is associated with general decline across a variety of cognitive domains, including memory, processing speed, executive functioning, and language production ability (Burke & Shafto, 2004, 2008; Salthouse, 2010). Life experiences, conceptual knowledge, and vocabulary however, tend to increase with age (Park et al., 2002; Verhaeghen, 2003). While some studies claim that age-enriched semantic information leads to greater interference (Buchler et al., 2007; Ramscar et al., 2017), other studies show that increased vocabulary is associated with lower error rates (Gollan & Goldrick, 2019) or with more lexically diverse output (Rabaglia & Salthouse, 2011), suggesting that vocabulary knowledge supports language production (Kavé & Halamish, 2015; Shafto et al., 2017). Moreover, there is evidence that semantic storage (i.e., vocabulary) positively interacts with other aspects of semantics, such as selection demands (i.e., executive functioning and language production (e.g., Burke & Shafto, 2008; Shafto et al., 2017). However, studies like these often do not focus on the overall structure of the lexicon.

Networks can be used to represent the structure of conceptual representations (Castro et al., 2020; Hills & Kenett, 2022; Siew et al., 2019). These computationally-modeled semantic memory networks allow the analysis of more than just word-word relations, but rather the interactions between one and many words, as well as overall structural characteristics of the network, such as efficiency. 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., 2018, 2019, 2022).

While these studies contributed seminal knowledge about the effect of age on semantic memory network structure, they largely focused on group level differences (but see Wulff et al., 2022) which precludes examining how language production abilities relate to network measures. Constructing semantic memory networks separately for each individual affords an opportunity to relate semantic memory network structure to individual differences in cognition. This individualized approach is particularly relevant in aging research as variability in cognitive performance increases with age and semantic knowledge is a function of personal experience. This leads to our primary research questions: are individually constructed semantic memory network measures related to language abilities, and are age-related differences in semantic memory networks related to age-related differences in vocabulary or language production?

The Aging Mental Lexicon

Cognitive aging is associated with declines in domain-general mechanisms including slower processing speeds, declines in working memory capacity, and decreased inhibitory control (Cabeza et al., 2018; Salthouse, 2010). Aging also affects specific cognitive processes such as language production, resulting in age-related difficulties in word retrieval (Burke & Shafto, 2008; Peelle, 2019). Though, these age-related word retrieval declines weaken in tasks where older adults have the opportunity to utilize their rich vocabularies (Gollan & Goldrick, 2019). Moreover, increased vocabulary knowledge can explain reduced error rates, faster reading times, and greater self-corrections for older adults (Gollan & Goldrick, 2019). Higher vocabulary scores have also been associated with fewer lexical retrieval difficulties (i.e., Tip of the Tongue states; TOTs) for adults in general, and larger vocabularies among older adults do not increase retrieval failures (Shafto et al., 2017). These findings are consistent with studies showing increased vocabulary knowledge allows older adults to produce more diverse and sophisticated discourse (Rabaglia & Salthouse, 2011) and more unique word associations (Burke & Peters, 1986). Burke and Peters (1986) concluded that verbal ability but not age, influenced semantic memory structure and semantic encoding.

Although older adults have larger vocabularies, a wealth of behavioral studies using a variety of tasks such as lexical decision, sentence and paragraph comprehension, and pronunciation tasks, have shown that older and younger adults have similar patterns of semantic priming, suggesting that spreading activation occurs similarly in younger and older adults (e.g., Howard 1981, 1983; Burke, White, & Peters, 1987; Laver & Burke, 1993; Madden et al., 1993; Balota & Duchek, 1988, 1989). Moreover, older adults produce similar word associations (Burke & Peters, 1987) and show similar frequency effects (e.g., Allen et al., 1993; Cohen-Shikora & Balota, 2016) compared to younger adults, again highlighting general similarities in older and younger adults’ semantic memory.

While the general pattern of semantic priming and word associations remains intact with age, older adults often have slower response times and activation may take longer to build (e.g., Balota & Duchek, 1988). In general, previous work has attributed differences in response time on semantic priming tasks to age-related slowing of sensory, motor, or decisional processes. Critically, these studies suggest that semantic retrieval mechanisms – like spreading activation – are preserved in older adulthood (Balota & Duchek, 1988; Madden et al., 1993). This complement of results (i.e., intact semantic priming patterns, coupled with slower responses) may be related to the role of automatic versus attentional aspects of semantic priming in relation to aging (e.g., Burke et al., 1987; Laver & Burke, 1993; Balota et al., 1992). Initially, a word may automatically yield a spread of activation through the semantic memory network reaching other semantically related words and increasing their resting activation level (Klimesch, 1987; Kroll & Klimesch, 1992). This automatic process benefits visual encoding and decreases response time to a subsequent target word. In addition to such automatic processes, individuals may also engage in more controlled processes (e.g., creating a list of expected associations), thereby increasing their level of activation through attentional processes. Further, these processes may interact as suggested by a semantic priming study by Balota and colleagues (1992). Overall, these two mechanisms (automatic and attentional spreading activation) can work together to boost word recognition (Laver & Burke, 1993), and highlight the influence of both automatic semantic processes, which remain intact with age (Balota et al., 1992), and reliance on other cognitive abilities, which may decline with age.

Semantic Memory Network Development Across the Lifespan

Some of the earliest studies on semantic memory were characterized in terms of spreading activation processes (Collins & Loftus, 1975) in which words are connected in terms of semantic similarity and activation spreads between closely related concepts (Balota & Duchek, 1988, 1991; Kenett et al., 2017; Kumar et al., 2020). Network science builds off these early network conceptualizations by providing a set of tools that mathematically quantify the characteristics of a system (Sporns, 2011). For example, measures like path length and global efficiency, which assess the exchange of information across the number of links (i.e., edges) between words, are conceptually similar to the concept of spreading activation. For example, previous research on semantic memory has shown that the path length in semantic networks corresponds to whether participants’ judge two concepts being related to each other (Kenett et al., 2017; Kumar, Balota, & Steyvers, 2020). Global efficiency of semantic memory networks has been related to individual differences in verbal creativity, which involves connecting weakly related concepts (He et al., 2020). Other network measures characterize the structure of the system. For example, clustering coefficient refers to the extent to which neighbors of a node will be neighbors of each other (i.e., a neighbor is a node i that is connected through an edge to node j). A higher clustering coefficient indicates a more interconnected semantic network (Siew et al., 2019). Modularity estimates how a network is partitioned into smaller communities (Fortunato, 2010; Newman, 2006). Such subcommunities can be thought of as subcategories in a semantic network (e.g., pets in the ‘animals’ category). Previous research has shown that modularity in semantic memory networks is inversely related to a network’s flexibility and that this measure typically increases with age (Cosgrove et al., 2021; Kenett et al., 2018, 2021).

Given the dynamic nature of semantic memory, studies have focused on network structure changes across development (Steyvers & Tenenbaum, 2005). This research suggests structural network changes both throughout child language acquisition (Hills et al., 2009a, Hills et al., 2009b) as well as during associative learning tasks in young adults (Mak & Twitchell, 2020). Since people continue to acquire knowledge and vocabulary across the lifespan, semantic memory structure may also continuously change with age (Bieth et al., 2021; Kenett & Thompson-Schill, 2020; Wulff et al., 2019; Yee & Thompson-Schill, 2016). One study assessed these lexical changes across the lifespan (i.e., 10–84 years) by mapping free association data from 8,000 participants (Dubossarsky et al., 2017). Free association tasks typically present participants with a target word and ask them to produce a continuous stream of thought associations for one minute (Dubossarsky et al., 2017). Early life networks were characterized by greater rigidity in structure, as measured by increases in average shortest path length, and lowered clustering coefficient, among other network measures (Dubossarsky et al., 2017). As age increased this pattern reverses – resulting in more sparsely connected networks. The authors concluded that the aging results are not an inverse of early life, but rather a reflection on the interaction between the accumulation of knowledge through life experiences, and declines in cognitive control processes that affect memory retrieval (Dubossarsky et al., 2017). Two recent studies used verbal fluency data to construct semantic networks, replicating and extending previous findings based on free association data (Cosgrove et al., 2021; Wulff et al., 2018). Overall, this work suggested that younger adults’ semantic memory networks have greater connections between words, less segregation of sub-communities, and greater flexibility for more efficient semantic processing (Cosgrove et al., 2021; Wulff et al., 2018). Critically, Wulff and colleagues’ simulation of aging networks suggested that a single underlying mechanism, likely deterioration, is responsible for the structural differences between younger and older adults (Wulff et al., 2018). Complementing these findings, Cosgrove and colleagues (2021) conducted a percolation analysis which quantifies the robustness of the semantic memory networks to simulated “attacks” (i.e., removing links in the network). They found that older adults’ semantic memory networks broke down faster compared to younger adults’ networks, suggesting diminished flexibility in older adults’ networks, despite the stability of semantic memory across the lifespan (Cosgrove et al., 2021). These age-related differences in the structure and robustness of semantic memory networks could then lead to the often-observed age-related differences in retrieval difficulty or tip-of-the-tongue states.

Individual Differences in Semantic Memory Network Structure

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 et al., 2018, 2019, 2022). In one such study, Wulff and colleagues calculated individual semantic networks using semantic similarity ratings (Wulff et al., 2018). Though these network measures were consistent with group findings (Cosgrove et al., 2021; Dubossarsky et al., 2017; Kenett et al., 2021; Wulff et al., 2019; Zortea et al., 2014), older adult semantic memory networks were not as generalizable across individuals as younger semantic memory networks. 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 et al., 2018, 2019). 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 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.

Current Study

The present study examines the underlying mechanisms of age-related differences in semantic memory. To build on previous group-level studies, we adopt an individual differences approach which better captures age-related variability, allowing a direct link between network measures and cognitive abilities. Previous research suggests that older adults’ semantic memory networks are more separated, more segregated, and more variable (Cosgrove et al., 2021; Dubossarsky et al., 2017; Kenett et al., 2021; Wulff et al., 2018, 2019). However, network research often relies on group averages, and it remains unclear if individual-based semantic network measures influence language production abilities related to aging. Prior work has shown that while individual-based semantic memory network differences are similar to group level findings, older adult networks were more variable and less generalizable across individuals (Wulff et al., 2018)—potentially due to greater variability in life experiences and acquired knowledge in older adults. Moreover, several studies that have examined individual-based network differences in older adults included relatively small samples (Wulff et al., 2022).

Increased semantic memory network variability with age suggests that individual differences may be a potentially important factor in cognitive aging, and previous work suggests that this approach is both useful and possible (Kenett & Faust, 2019a; Wulff et al., 2019; Zemla & Austerweil, 2018). Here we assessed individual-based semantic memory networks via a semantic relatedness judgment task (RJT). Performance on semantic judgment tasks are much more strongly influenced by meaning as the decision relies on semantic aspects of the stimuli (Pexman et al., 2017; Pexman & Yap, 2018; Yap et al., 2012). Such relatedness judgments can be used as a proxy for how concepts are organized and has been applied across different languages and cultures (Benedek et al., 2017; He et al., 2020; Ovando-Tellez et al., 2022). Thus, the RJT is well-suited to studying individual differences in the aging lexicon (Wulff et al., 2022).

In addition, calculating individually based semantic memory networks allows us to compare the structural network measures with each individual’s language ability. In order to assess both language production and representation, we used verbal fluency and vocabulary tasks. The verbal fluency task measures an individual’s production performance under a given time constraint, while the untimed vocabulary assessment reflects their semantic store. We hypothesized that individual differences in semantic network properties, as measured via the RJT, will mimic group level findings, and would be mediated by age and language abilities. Specifically, older adults’ semantic memory networks should be more segregated, less connected, and less efficient compared to younger individuals, and these age-related differences would weaken with better vocabulary knowledge and verbal fluency.

Methods

Transparency and Openness

The de-identified data on which this study’s results and conclusions are based are openly accessible. The study design, hypotheses, and analytic plan were not preregistered.

Participants

Twenty-eight younger adults (18 – 30 years, M = 23.18 years old, SD = 3.93, females = 11, males = 15, non-binary = 1, and 1 individual who preferred not to answer) and twenty-eight older adults (60 – 80 years, M = 64.96 years old, SD = 3.94, females = 19, males = 9) were recruited online through Prolific (https://prolific.co/). No information about race was collected for this online study. Based on a post-hoc power analysis, as well as a prior aging and semantic memory network study (Cosgrove et al., 2021), 56 participants should provide greater than 80% power to detect a large sized effect in semantic memory organization between the two age groups (p < .05). All participants reported no history of neurological disorder or disease and were fluent English speakers. All participants who chose to take part in these studies provided informed consent and were paid $10/hr for their participation. All experimental procedures were approved by the Pennsylvania State University Institutional Review Board as part of the Online Studies of Language Protocol STUDY00014804.

Verbal fluency task

Participants completed two commonly used categorical verbal fluency tasks – animals and fruits & vegetables categories. For each of these categories, participants were presented with the category (animals or fruits & vegetables) and were given one minute to type as many members of each category that they could think of (Benton, 1968). Performance on the verbal fluency task was calculated through the number of category correct responses. For our analyses we combined the two categories since the two tasks were highly correlated, r = .74, p < .001.

Vocabulary Knowledge Assessment

Participants also completed a vocabulary assessment which included two parts from the ETS Kit of Factor-Referenced Cognitive Tasks: the Advanced Vocabulary Test II (18 items, 4 minutes) and the Extended Range Vocabulary Test (24 items, 6 minutes; Ekstrom & Harman, 1976). Both tests included multiple choice questions that presented a target word with four to five answer choices. Participants were asked to choose the word that best described the target word. Performance on these tasks was measured as the sum of overall accurate responses and these tasks were combined for analyses, r = .71, p < .001. In addition to this crystallized intelligence measure, a fluid intelligence score was collected through the Culture Fair Intelligence Test (CFIT; Cattell & Cattell, 1960).

Relatedness Judgement Task

Semantic memory networks are represented by concepts (i.e., nodes) and the semantic relatedness between them (i.e., edges). Semantic relatedness judgements are representative of the semantic distance between concepts in a participant’s semantic memory network (Kenett et al., 2017). Using a similar methodology to Benedek and colleagues (2017), participants judged the relatedness of all possible pair combinations from a list of 28 words (See Appendix A for the word list). The stimuli were chosen across 7 categories and are all considered highly frequent, concrete, and have an early age of acquisition (Linguistic features of the words are provided in Appendix B). The stimuli were chosen across several categories so that they would correlate more with within category than across category. There were no differences in lexical characteristics between each category (see Appendix C). In addition, these categories (e.g., animals, furniture, tools, etc.) are relatively stable across age and cohort effects as described in Castro et al., 2021.

The order of the word-pairs for each participant was randomized (378 pairs). The word pairs were presented in the center of the screen with a continuous scale from 0–1 below the words. The scale had a slider with ends labeled “unrelated” and “strongly related”. On each trial, the slider was initially positioned in the middle of the scale (See Appendix D for an example display). Participants were given the instruction to provide a quick, instinctive judgement for the semantic relatedness of the word pair and to avoid over-analyzing the relationship. Participants judged each unique pair once. The judgement units were measured by a fraction of the segment’s width. Data was rounded to the tenth decimal place (i.e., 0.1) before computing the semantic networks.

Network Estimation

For each participant, their semantic relatedness judgements were converted into a 28 × 28 adjacency matrix of a weighted, undirected semantic memory network (Benedek et al., 2017). Individual-based semantic memory networks were estimated for each participant. In these estimated weighted undirected networks (WUN), all edges were kept in the network and edges are weighted based on the judgements provided by each participant during the task (Benedek et al., 2017). The benefit of this methodology is that it avoids any arbitrary thresholding of edges for network filtering. This is critical for capturing the possible weaker connections of semantic relationships in aging lexicons (He et al., 2020; Ovando-Tellez et al., 2022). The relatedness judgement task allows us to create individual correlation matrices that are used to calculate the semantic networks. Previous aggregated networks combined participant responses to estimate one semantic memory network per age group where there were only two sets of network measures for comparison (Cosgrove et al., 2021). Our network metrics (e.g., modularity, clustering coefficient, and global efficiency) are thus tailored to each individual. This, then allowed us to compare these network measures to individual differences in vocabulary knowledge, verbal fluency performance, and fluid intelligence.

Semantic Memory Network Measures

Analyses for the individual-based semantic memory network measures were calculated using the Brain Connectivity Toolbox for MATLAB (Rubinov & Sporns, 2010). For each individual’s semantic memory network, we examined the clustering coefficient, global efficiency, and modularity (Newman, 2006; Sporns, 2011). We chose these network metrics as they examine multiple levels of semantic memory organization and have been used in previous aging network studies (e.g., node to node relationships, across entire network relationships, Benedek et al., 2017; He et al., 2020; Ovando-Tellez et al., 2022). Clustering Coefficient (CC) represents the connectedness of the neighbors of a node, and it influences the flow of information through and around a node; more connections indicate a more robust network (Sporns, 2011). Global Efficiency (EGlobal) takes the inverse of path length – or the relatedness between any pair of nodes, averaged across the network (Sporns, 2011). Shorter path lengths indicate higher levels of efficiency and therefore the faster information can travel across the network. Finally, modularity (Q) is the community or grouping structure of nodes in the network (Fortunato, 2010; Newman, 2006).

Analyses

We ran independent sample t-tests to examine whether the individual-based semantic memory network measures of CC, EGlobal, and Q differed between younger and older adults. Then, collapsing across age groups, individual-based semantic memory network measures of CC, EGlobal, and Q were compared to verbal fluency performance and vocabulary knowledge via linear regression analysis. In addition, we ran mediation analyses to assess whether including verbal fluency and vocabulary performance affected the age effect on the individual-based semantic memory network measures of CC, EGlobal, and Q. A few missing data points in the vocabulary extended task and the animals verbal fluency task were imputed using the mice package in R (van Buren et al., 2011). The mice package imputes missing data with plausible values calculated through stochastic regression imputation with the complementing verbal fluency and vocabulary tasks. R script, regression graphics, and imputed data can be found on OSF (https://osf.io/rh76p/).

Results

The average semantic relatedness between 28 concepts was 0.30 (SD = 0.13) for younger adults and 0.18 (SD = 0.11) for older adults. Averaged across responses, younger adults judged the concepts as more related than older adults t(56)= −3.94, p < .001. The standard deviation and range of these judgement were not significantly different between age groups. The semantic judgments with values of 0 were adjusted to be .01 because zero values disproportionately affected some of the mathematical calculations.

To examine the relationship between individual-based semantic memory network properties and language, individual-based semantic memory network measures of CC, EGlobal, and Q were correlated with performance on a verbal fluency task and vocabulary knowledge. An example of a younger and older adult semantic network is shown below (Figure 1). See Appendix E for a visualization of the individual differences between all 56 semantic memory network structures of younger and older individuals. In addition, correlation matrices across all measures are reported separately for the younger and older adults in Appendix F.

Figure 1.

Examples of an individual-based semantic memory network for a younger adult (left) and older adult (right) calculated from their semantic relatedness judgements.

Similar to group level differences that have been observed by others (Cosgrove et al., 2021; Dubossarsky et al., 2017), significant age effects were found between the individual-based semantic memory network measures (Figure 2). Specifically, through independent sample t-tests, we found that clustering coefficient was lower for older adults, t(54) = −2.07, p = .04, global efficiency was lower for older adults, t(54) = −2.03, p = .047, and modularity was higher for older adults, t(54) = 5.06, p < .001. Overall, the within-group variance for each network measure was greater for older adults compared to younger individuals (Table 1).

Figure 2.

Differences among age and semantic network measures of clustering coefficient (left), global efficiency (middle), and modularity (right). Older adults’ semantic networks exhibited lower clustering coefficients and global efficiency values, yet greater modularity properties compared to younger adults.

Table 1.

Summary statistics for network measures of clustering coefficient (CC), global efficiency (Eglobal), and modularity (Q), as well as verbal fluency and vocabulary performance for younger and older adults.

	Younger Adults	Older Adults
CC Mean (SD)	0.31 (0.11)	0.24 (0.11)
CC Range	0.13 – 0.51	0.06 – 0.53
Eglobal Mean (SD)	0.46 (0.10)	0.38 (0.16)
Eglobal Range	0.17 – 0.66	0.04 – 0.70
Q Mean (SD)	0.16 (0.10)	0.31 (0.12)
Q Range	0.04 – 0.45	0.13 – 0.50
Verbal Fluency Total Mean (SD)	37.29 (15.05)	35.68 (14.51)
Verbal Fluency Total Range	11 – 71	8 – 63
Vocabulary Total Mean (SD)	20.94 (5.30)	28.54 (7.72)
Vocabulary Total Range	11 – 30	5 – 38

Consistent with the aging literature, vocabulary knowledge was significantly greater for older adults, t(54) = 4.29, p < .001. In addition, after a Levene’s test indicated homogeneity of variances between the two age groups on the fluid intelligence measure, (F = .26, p = . 61), an independent sample t-tests showed that fluid intelligence scores were lower for older adults, t(53) = −2.9, p = .005. Next, we examined the relation between the individual-based semantic memory network measures and performance in the language tasks. Vocabulary was significantly related to the individual-based semantic memory network measures (Figure 3). Mimicking the age effects, clustering coefficient, b = −0.005, t(54) = −3.13, p = .002, and global efficiency, b = −0.006, t(54) = −2.64, p = .01, had a negative relationship with vocabulary; while modularity, b = 0.006, t(54) = 2.87, p = .006, was positively related to vocabulary knowledge. This finding indicates that vocabulary knowledge had a significant effect on semantic memory network organization as measured by the RJT. However, these individual-based semantic memory network measures of clustering coefficient (p = .95), global efficiency (p = .41), and modularity (p = .71), did not significantly correlate with language production ability as measured by the total number of verbal fluency responses. There was also no significant difference in the total number of verbal fluency responses produced across younger and older adults (p = .69).

Figure 3.

Effect of vocabulary knowledge on individual-based semantic memory network measures of clustering coefficient (left), global efficiency (middle), and modularity (right). Xaxis – vocabulary knowledge. Y-axis – network measure. Shaded gray area represents standard deviation. Green data points represent younger adults, orange data points represent older adults.

In addition, a mediation analysis was run using the mediation package in R (Tingley et al., 2014), to quantitively measure whether vocabulary knowledge mediated these age effects. Adding vocabulary performance to the mediation analysis did not significantly change the effect of age on the individual-based semantic memory network measures of EGlobal and Q. The effect of age on clustering coefficient was partially mediated by vocabulary knowledge. The regression coefficient between age group and clustering coefficient and the regression coefficient between vocabulary and clustering coefficient was significant (Figure 4). The indirect effect was (0.87)*(0.3) = −0.26. We tested the significance of this indirect effect using bootstrapping procedures. Standardized indirect effects were computed for each of 500 bootstrapped samples, and the 95% confidence interval was computed by determining the indirect effects at the 2.5th and 97.5th percentiles. The bootstrapped standardized indirect effect was −0.26, and the 95% confidence interval ranged from −0.593 to −0.07. Thus, the indirect effect was statistically significant (p=.008). Verbal fluency did not mediate associations between age and network measures (see Table 2 for more details). Since we did not find a direct effect of age and verbal fluency responses, this result is not surprising.

Figure 4.

Mediation analysis with age group (older adults = 1, younger adults = 0) as the independent variable, vocabulary knowledge as the mediator, and individual-based semantic memory network measures of clustering coefficient (left), global efficiency (middle), and modularity (right) as the dependent variables. Path coefficients are standardized β-coefficients from (multiple) regression analyses. The coefficient below the dotted line is the standardized β-coefficients for the mediated effect.

Table 2.

Mediation analysis for vocabulary and verbal fluency performance.

Mediation Model	Beta Value	p value
Vocabulary mediating Age—Modularity Relationship	0.1095	0.22
Vocabulary mediating Age—Clustering Coefficient Relationship	−0.258	0.008*
Vocabulary mediating Age—Global Efficiency Relationship	−0.207	0.068
Verbal Fluency mediating Age—Modularity Relationship	0.00181	0.83
Verbal Fluency mediating Age—Clustering Coefficient Relationship	−0.0001	0.772
Verbal Fluency mediating Age—Global Efficiency Relationship	−0.00251	0.884

Discussion

In the current study we examined the relation of individuals’ semantic memory networks, vocabulary knowledge and language production, and how these relations vary across the lifespan. The application of computational network science methodologies to cognition and aging allows us to investigate these issues, advancing our understanding of the effect of continued knowledge acquisition throughout the lifespan on individuals and how variations in such processes relate to changes in ones’ semantic memory structure.

Semantic cognition comprises both overall conceptual knowledge, the control processes that influence selection, and the interaction between these two mechanisms (Lambon Ralph et al., 2017). Critically, how our semantic memory is organized can influence this interaction. The emerging field of cognitive network science provides mathematical tools to quantify the structural properties of semantic memory through network measures (Castro & Siew, 2020; Hills & Kenett, 2022; Siew et al., 2019). These measures can be used to examine how words are organized and retrieved both at a global network level as well as at a detailed word level (Siew et al., 2019). The intersection of aging and semantics provides an excellent case for examining differences in network structure, as vocabularies grow throughout adulthood, thereby changing semantic memory organization (e.g., Burke & Shafto, 2008). Moreover, cognitive network science research has shown age-related structural differences, suggesting that semantic memory organization is less efficient, more segregated, and less interconnected in older adults, compared to younger adults (Cosgrove et al., 2021; Dubossarsky et al., 2017; Kenett et al., 2021; Wulff et al., 2018, 2019, 2022). Moreover, assessing individual differences in semantic memory structure is particularly important among older adults who often show increased variability. Yet, among the aging literature, the majority of studies have examined group level differences (Wulff et al., 2019), raising the question of whether individual variation in relevant language abilities, such as vocabulary knowledge and verbal fluency, impacts the structure of aging semantic networks.

We found that individual network metrics significantly differed between younger and older adults. In line with previous work, the semantic memory networks of older adults were less efficient (lower global efficiency), less interconnected (lower clustering coefficient), and more segregated (higher modularity) compared to younger adults. Interestingly, there was no relationship between semantic memory network measures and verbal fluency performance, but vocabulary knowledge significantly predicted semantic memory network measures, suggesting that the ability to fluently retrieve knowledge from semantic memory is less relevant than the amount of knowledge in memory for explaining age-related semantic network change.

Vocabulary knowledge significantly mediated the effect of age on clustering coefficient – a measure of interconnectedness among words in the semantic memory network – but did not affect the other semantic memory network measures (i.e., global efficiency and modularity). While these network measures are related to one another, some key differences remain. Clustering coefficient measures the connectivity or richness of semantic neighborhoods. Whereas modularity measures two different factors – how much the network breaks apart into sub communities, and how much these subcommunities have higher within-group clustering coefficient and lower between-group connectivity. In other words, our results suggest that a person’s vocabulary size may influence how their words are connected in their semantic memory network but not how these concepts cluster together into subcategories or their overall network efficiency. The absence of an effect of vocabulary on processing efficiency is supported by previous behavioral research where age-related increases in vocabulary knowledge did not increase Tip of the Tongue states (Shafto et al., 2017). Since vocabulary knowledge increases with age, we conclude that 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.

Aging Semantic Memory Networks

Changes in semantic memory network structure as a result of healthy aging could occur because of differences in the size and structure of the network (Wulff et al., 2018, 2019). Several studies suggest that as people age their semantic memory networks become more organized – represented by longer path lengths, greater modularity, and lower clustering coefficients (Cosgrove et al., 2021; Dubossarsky et al., 2017; Wulff et al., 2018). Through a percolation analysis (i.e., targeted attacks), we previously showed that this greater organization has a cost in relation to network flexibility (Cosgrove et al., 2021). Vocabulary knowledge increases with age, and larger semantic memory networks could hinder search processes, contributing to the memory declines seen in the aging literature. These behavioral changes likely lead to variation in semantic memory network structure between older and younger individuals. The current study investigates the role of vocabulary knowledge in these age-related semantic network differences. We find that as individuals continue to acquire knowledge, this enriched vocabulary leads to maturation in the semantic memory network. As we expected, the results of this study showed that older adults have more segregated and less efficient semantic memory networks exhibited by higher modularity and lower global efficiency values. Here, we replicate and extend our previous study’s findings using a different semantic task to construct individual-based semantic memory networks, demonstrating the robustness of our results across two different tasks.

Previous studies have emphasized the importance of assessing individual differences in the underlying semantic memory network structure relevant for other cognitive abilities, such as memory retrieval and creative thinking (Kenett & Faust, 2019a; Wulff et al., 2019; Zemla & Austerweil, 2018). Wulff and colleagues (2018), for example, estimated semantic memory networks from verbal fluency responses of younger and older adults from three separate studies. Semantic memory networks of older adults were found to have fewer connections, were less robust, and less efficient compared to younger adults – consistent with other network studies on aging (Cosgrove et al., 2021; Dubossarsky et al., 2017; Kenett et al. 2021; Wulff et al., 2019). Extending the aggregated results of age-related differences in network structure, individual differences were calculated from semantic similarity ratings (Wulff et al., 2018). Critically, the similarity ratings used for older adult semantic memory networks were not as generalizable across individuals as for younger semantic memory networks. Our results align with the Wulff et al. (2018) study such that the variability across the three network measures increased among our older adult group (see Table 1). Thus, our results are consistent with prior aging literature using both individual differences and group averages approaches, where older adults’ semantic memory networks are characterized by longer path lengths, greater modularity, and lower clustering coefficients – indicating a less efficient and less organized network. These structural differences may be due to a variety of factors that may not necessarily reflect cognitive decline but rather differences in continued learning across the lifespan (Wulff et al., 2018).

Individual Differences in Semantic Memory Representation

Life experiences and conceptual knowledge increase across the lifespan: older adults have larger vocabularies and are more confident in their word knowledge (Verhaeghen, 2003; Kavé & Halamish, 2015). Although there are well documented age-related differences in vocabulary, semantic knowledge does not exist in a vacuum. Some research has claimed that older adults have different memory representations due to the fact that they are exposed to more environmental input (Ramscar et al., 2017). Further, lifelong word and association learning has accounted for performance declines in older adults compared to younger adults across a variety of tasks including recognition and word pair learning (Buchler et al., 2007; Ramscar et al., 2017, respectively). Several studies utilizing picture naming and lexical retrieval tasks have shown an interaction between semantic storage and semantic selection (Burke et al., 1991; Lambon Ralph et al., 2017; Shafto et al., 2017). These results, however, do not directly link increased knowledge with interference on word retrieval. Rather, they provide additional evidence that crystallized intelligence supports successful word retrieval, but this relationship weakens with age (Shafto et al., 2017). While age-related deficits in single word retrieval are consistent across tasks, enhanced semantic context, from larger vocabularies, can help older adults during discourse production (Kavé & Goral, 2017). These findings provide further support for an adaptable semantic memory network structure, dependent on the accumulation of knowledge and individual variability in life experiences.

While our individual-based semantic memory network measures largely replicate previous aggregated results in the aging literature (Cosgrove et al., 2021; Dubossarsky et al., 2017; Kenett et al., 2021; Wulff et al., 2019)(Wulff et al., 2022), more evidence is necessary to substantiate any claims that group-based semantic memory networks are generalizable across individuals – especially in aging populations. Moreover, it is important to evaluate these results across multiple linguistic tasks, as highlighted by our results for verbal fluency and vocabulary knowledge.

Though non-linguistic cognitive abilities, such as speed, working memory, and inhibition decline with age, and these general cognitive changes can affect language ability (Caplan & Waters, 2005; Hasher et al., 2008; Hoffman et al., 2018; Kemper et al., 2003a, 2003b, 2010; Pickering & Garrod, 2013; Salthouse, 2010), these mechanisms may not directly affect the results of this study. The relatedness judgement task used to estimate semantic networks was untimed. Previous studies have shown that activation processes are relatively preserved with age, particularly when untimed measures are incorporated (Balota & Duchek, 1988; Burke & Shafto, 2008). Additionally, we did not find an age difference in our verbal fluency production data – a task that reflects both language and executive functioning abilities. While the Inhibitory Deficit Theory argues that the active, goal-directed process of inhibition is the main source of variation in cognitive performance with age (Hasher et al., 2008), age-related differences in inhibition are unlikely to be driving our results because we did not observe age-related differences in a language production task that involves executive function (i.e., verbal fluency).

Limitations and Future Directions

In this study, we compared our network results to verbal fluency data and found no significant relationship. In addition to finding no significant age effects in the total number of verbal fluency responses, we also did not find any age differences in clustering or switching behaviors. While the verbal fluency and semantic relatedness judgment tasks both tap into aspects of semantic processing, it is possible that the underlying cognitive mechanisms required for each task are different. A semantic verbal fluency task focuses on a single category (e.g., animals) and instructs the participants to produce as many words belonging to this category in one minute. This time sensitive task utilizes semantic control mechanisms, including aspects of executive functioning, to remain focused and inhibit any incorrect responses (Amunts et al., 2020). Consistent with this, an analysis of younger adults showed that semantic verbal fluency strongly correlated with executive factors of working memory, fluid reasoning, and mental flexibility (Aita et al., 2018). In their study, individuals who performed well on the verbal fluency task also had better executive functioning ability, substantiating the claim that verbal fluency should be considered an executive language task (Aita et al., 2018). Alternatively, the semantic memory networks in this study were created from the relatedness judgment task, where participants judge how related two concepts are to one another. In this task participants are likely employing more representational aspects of semantics, highlighting the semantic store. Although the present experiment was sufficiently powered to detect a large effect size (d > .8), it could be the case that the effect size related to language production and semantic networks is smaller, and a significant effect would be found with a larger sample size.

Another limitation of this study is that our approach was not able to capture all relevant individual differences in semantic representations given that we included a relatively small number of highly frequent and concrete words. This, however, is a common approach and a common limitation due to time constraints. These stimuli have been used successfully in the past to create semantic memory networks across several studies, labs, and languages (Benedek et al., 2017; He et al., 2020; Ovando-Tellez et al., 2022). Prior work in the aging literature has shown that during picture naming word retrieval, older adults have lower accuracy for low frequency items compared to younger adults (Burke et al., 1991; Ferre et al., 2020; Gertel et al., 2020), which suggests that even larger age-related differences may be found with less frequent words. Future studies should expand the breadth of cue words to include lower frequency items that may lie along the periphery of the semantic network and may reveal greater age-related differences.

Another limitation of the present work relates to general sampling issues in aging research. Online data collection is beneficial in cognitive aging research – allowing scientists to capture a more diverse representation of participants as well as assessing whether age-related effects generalize to non-laboratory environments (i.e., improving ecological validity of the results). An open question remains, however, about the sampling of older adults who participate in these online studies and whether their data is a reliable subset of the larger population (Greene & Naveh-Benjamin, 2022). While we did not collect details on educational attainment or occupation, we did collect a fluid intelligence score from the Culture Fair Intelligence Test (CFIT; Cattell & Cattell, 1960). An independent sample t-tests showed that fluid intelligence scores were lower for older adults, t(53) = −2.9, p = .005. These findings are consistent with the aging literature (e.g., Park et al., 2002) and suggest that our older adult sample reflects typical cognitive aging.

In addition, although we assessed individual variability between younger and older adults, our data consisted of an extreme age group comparison. Including a middle-aged participant group would allow for a more continuous lifespan sample to better measure individual variability in the relationship between language processing (i.e., production and vocabulary), semantic memory network properties, and age. Moreover, despite its utility, the cross-sectional approach incorporates only a single time point measured across individuals. This means that group differences may also include cohort effects and impair our ability to assess aging effects. This concern also relates to the mediation analyses (e.g., Hofer & Sliwinski, 2001; O’Laughlin et al., 2018). For this reason, we interpret the results as suggesting that vocabulary and semantic network structure are related. The causal nature of this relationship cannot be fully examined without longitudinal research and alternative causal chains where a third variable explains age-related differences in both vocabulary and semantic network structure is also possible (e.g., Lindenberger et al., 2011; Raz & Lindenberger, 2011). While our findings shed light on the relationship between individual network structure, age, and vocabulary knowledge, the practical consequences of age-related differences in network structure are still emerging and this remains a promising area for future research.

Conclusions

The present study examined individual differences in semantic memory network structure across the lifespan. Overall, the estimated individual-based semantic memory networks of younger and older adults largely replicated prior findings from aggregated semantic memory networks: older adults’ semantic memory networks showed lower levels of efficiency, but increased modular structure (Cosgrove et al., 2021; Dubossarsky et al., 2017; Kenett et al., 2021; Wulff et al., 2019; Wulff et al., 2022). Extending previous work, we found that these semantic memory network differences reflect age-related increases in vocabulary knowledge, which are consistently found across the aging literature (Park et al., 2002; Verhaeghen, 2003). Previous work has associated these age differences to the deterioration of links in semantic memory networks (Wulff et al., 2018). However, our findings show that age-related increases in vocabulary knowledge reflect the critical role that the accumulation of knowledge has on the structure of semantic memory.

Public Significance Statement:

Although knowledge and vocabulary continue to grow throughout the lifespan, this expansion affects how information is stored in semantic memory. Computational network analysis showed that increases in vocabulary were associated with semantic memory networks that were less interconnected, less efficient, and more modular.

Appendix

Appendix A:

Stimulus words used in the semantic relatedness judgment task (Benedek et al., 2017)

Animals	Nature	Food	Tools	Furniture	Clothes	Container
Dog	Mountain	Meat	Hammer	Bed	Shoe	Bucket
Cat	Tree	Cheese	Broom	Table	Trouser	Bottle
Bird	River	Water	Pen	Cupboard	Hat	Sack
Cow	Sun	Bread	Knife	Lamp	Umbrella	Can

Appendix B:

Linguistic features of stimulus words in the semantic relatedness judgment task defined by the English Lexicon Project (Balota et al., 2007) and Brysabert et al., 2014 concreteness ratings.

Word	Length	Freq_HAL	Log_Freq_HAL	Concreteness	Age of Acquisition
knife	5	7,120	8.871	4.9	4.15
water	5	105,961	11.571	5	2.37
umbrella	8	1,930	7.565	5	5.68
trouser	7	186	5.226	4.89	7.22
tree	4	27,218	10.212	5	3.57
table	5	56,081	10.935	4.9	4.39
sun	3	74,083	11.213	4.83	3.4
shoe	4	5,209	8.558	4.97	2.6
sack	4	2,220	7.705	4.84	5.83
river	5	27,274	10.214	4.89	4.9
pen	3	7,902	8.975	4.92	5.11
mountain	8	34,727	10.455	4.96	6.15
meat	4	14,861	9.606	4.9	4.42
lamp	4	6,657	8.803	4.97	4
bed	3	31,345	10.353	5	2.89
hat	3	11,748	9.371	4.88	3.33
hammer	6	6,714	8.812	4.77	5.42
dog	3	58,314	10.974	4.85	2.8
cupboard	8	564	6.335	4.79	6.39
cow	3	7,262	8.89	4.96	3.94
cheese	6	8,662	9.067	4.7	4.33
cat	3	38,649	10.562	4.86	3.68
can	3	1,625,073	14.301	4.55	4.32
bucket	6	3,329	8.11	4.96	5.61
broom	5	600	6.397	4.89	5.5
bread	5	9,063	9.112	4.92	3.58
bottle	6	18,633	9.833	4.91	3.56
bird	4	19,070	9.856	5	3.52

Appendix C:

Multiple Group Comparisons of Lexical Characteristics between the 7 Categories of Stimulus Words yields no significant differences.

Lexical Characteristic	F-Value	P-Value
Length	2.52	0.1
Log Frequency	2.45	0.11
Concreteness Rating	0.62	0.71
Age of Acquisition	2.33	0.12
Semantic Neighborhood Density	0.71	0.65

Appendix D:

Example of the Relatedness Judgement Task Display.

Appendix E:

An example visualization of individual semantic memory networks estimated from the Relatedness Judgement Task (RJT) for the 28 younger (top) and 28 older adults (bottom).

Appendix F:

Correlation matrices across all measures are reported separately for the younger (top) and older adults (bottom).