Two pathways in vocabulary development: Large-scale differences in noun and verb semantic structure

Highlights

• Noun and verb semantic knowledge interact in adults.

• Noun and verb knowledge may impact noun and verb learning in children.

• Early noun and verb semantic structure was measured using tools of network science.

• Early noun knowledge supported later noun and verb learning.

• Early verb knowledge had different relationships with noun and verb learning.

Abstract

In adults, nouns and verbs have varied and multilevel semantic interrelationships. In children, evidence suggests that nouns and verbs also have semantic interrelationships, though the timing of the emergence of these relationships and their precise impact on later noun and verb learning are not clear. In this work, we ask whether noun and verb semantic knowledge in 16–30-month-old children tend to be semantically isolated from one another or semantically interacting from the onset of vocabulary development. Early word learning patterns were quantified using network science. We measured the semantic network structure for nouns and verbs in 3,804 16–30-month-old children at several levels of granularity using a large, open dataset of vocabulary checklist data. In a cross-sectional approach in Experiment 1, early nouns and verbs exhibited stronger network relationships with other nouns and verbs than expected across multiple network levels. Using a longitudinal approach in Experiment 2, we examined patterns of normative vocabulary development over time. Initial noun and verb learning was supported by strong semantic connections to other nouns, whereas later-learned words exhibited strong connections to verbs. Overall, these two experiments suggest that nouns and verbs demonstrate early semantic interactions and that these interactions impact later word learning. Early verb and noun learning is affected by the emergence of noun and verb semantic networks during early lexical development.

Introduction

In the adult language system, nouns and verbs have varied and multilevel semantic interrelationships. Noun and verb vocabulary knowledge is often conceptualized as being structured along associative pathways and taxonomic hierarchies (Fellbaum, 1998, Levin, 1993, Schuler, 2005). Evidence for semantic relations among nouns extends beyond lexicographical taxonomies of meaning – studies of lexical processing in adults show that nouns prime other nouns that have similar, thematically linked, or associated meanings (e.g., Hoey, 2012, Jones and Estes, 2012). And just as with nouns, verbs prime semantically related and associated verbs (e.g., Gomes et al., 1997).

Noun and verb semantics also demonstrate pervasive interactions with each other; adults’ knowledge of verbs activates associated nouns and their knowledge of nouns activates associated verbs. For example, in sentence processing contexts, adults use their knowledge of the semantics of verbs alone (e.g., Altmann & Kamide, 1999) and nouns and verbs together (e.g., Kamide et al., 2003) to predict upcoming nouns. Other work shows similar interactions among noun and verb semantics, including in priming tasks (Ferretti et al., 2001, McRae et al., 2005) and word description and rating tasks (McRae et al., 1997).

How do adults come to possess this complexly structured noun and verb vocabulary knowledge? Studies of language development in young children suggest that these semantic relationships may form early. For example, semantic relationships between nouns affect early noun processing and learning (Arias-Trejo and Plunkett, 2009, Arias-Trejo and Plunkett, 2013, Peters and Borovsky, 2019) and there is also evidence for similar effects within verb vocabularies (Saji et al., 2011). In some cases, semantic relationships facilitate word processing and learning; for example, Arias-Trejo & Plunkett (2009) found that 21-month-olds were faster to look at an object after having heard a semantically related noun, consistent with behavior seen in priming studies with adults. In other cases, semantic relationships complicate or slow word processing and learning, as in Saji et al. (2011), which found that verbs with similar meanings required an extended period of time for children to learn.

Children's noun and verb knowledge may also be interlinked across parts of speech. Children use a variety of information to develop early conceptions of lexical category to understand that nouns form a grammatical class distinct from verbs, including word cooccurrence (Mintz, 2003), sound-meaning correspondence (Dingemanse et al., 2015), and morphology (Onnis & Christiansen, 2008), raising the possibility that children’s noun and verb knowledge may be distinct in early vocabularies and influence each other. Evidence suggests that noun semantic knowledge indeed can support verb learning and processing in children. For example, the perceptual and semantic properties of the noun participants in events influence verb learning (Childers et al., 2016, Haryu et al., 2011, Kueser et al., 2023), indicating that children, like adults, pay attention to the semantic characteristics of nouns and use this information to support verb processing. Other work shows that familiarity with nouns and use of semantically descriptive nouns in verb-learning contexts help verb learning in children (Arunachalam and Waxman, 2011, Arunachalam and Waxman, 2015, Gleitman et al., 2005, Kersten and Smith, 2002).

Verb knowledge also appears to support noun learning and processing in children. Like adults, toddlers use verb knowledge to predict following nouns in sentence processing tasks (e.g., Fernald et al., 2008, Mani and Huettig, 2012) and can even combine noun and verb semantics to do so (Borovsky et al., 2012). In addition, toddlers can use knowledge of verbs’ semantics to learn novel nouns (Lidz et al., 2017, White and Lidz, 2022) and associations between nouns and novel verbs are readily learned (Yuan et al., 2011).

However, much remains to learned about how words’ semantic linkages mature across development, both within noun and verb vocabularies and between these parts of speech. One possibility that we will term the “Early Isolation hypothesis” is that semantic knowledge associated with individual words is initially isolated and separate, with semantic relationships between words only developing over time. Children’s knowledge of a verb like “eat”, for instance, may be initially separate from their knowledge of a noun like “cookie” and associative, taxonomic, thematic, and other semantic links between the words would take time to develop. The likelihood of this hypothesis being true may differ across parts of speech; that is, semantic relationships between nouns, verbs, and between nouns and verbs may develop at different times in the course of vocabulary development.

Some evidence suggests that the Early Isolation hypothesis is reasonable, particularly very early in vocabulary development for some word relationships. The idea that certain kinds of semantic information or associations take time to “come online” is supported by developmental findings of delayed onset of noun-noun priming through taxonomic (Arias-Trejo and Plunkett, 2013, Rämä et al., 2013) and other associations (Arias-Trejo & Plunkett, 2009) and varying degrees of contribution of higher-level semantic information to noun vocabulary growth (Peters & Borovsky, 2019). The evidence for the Early Isolation hypothesis is also suggestive for verb-verb and noun-verb relationships. The complexities associated with verb meanings may make semantic relationships between them more difficult to establish and delay their development. Given that early understandings of verb meanings may be imprecise (Saji et al., 2011) and that even adults have difficulty guessing which verb was used to refer to an action during play (Gillette et al., 1999, Zhang et al., 2022), the semantic boundaries between verbs are likely fuzzy and flexible throughout life, potentially complicating the formation of semantic links between verbs in children’s vocabularies. Actions are also more transient and less imageable than objects (Bird et al., 2001), meaning that it may be difficult for a child to extract the action-related information from a complex event to make fine-grained semantic comparisons between a novel action and a known one; verb-verb semantic relationships may then remain unnoticed by children until more advanced event processing skills develop. In addition, instances of verb usage in the language input can be marked by considerable semantic-syntactic variation (Talmy, 1985); for example, “give” can appear with one (“give it”), two (“give it to me”), or three arguments (“you give it to me”), which again may make consistent verb-verb semantic linkages difficult to establish.

Noun-verb relationships may also take time to develop because verbs tend to be learned later and in smaller quantities than nouns by young children in a variety of languages (Au et al., 1994, Frank et al., 2021). This temporal lag between noun and verb acquisition could induce separated and non-interacting noun–verb semantics in early vocabulary development. Last, evidence suggests that nouns and verbs – or more generally, words with more concrete vs. abstract meanings – are processed in distinct brain regions, a finding that suggests that neural connections between these regions may take time to mature (Vigliocco et al., 2011). Together, these differences may cause noun and verb semantics to be initially structured in different ways through different word-learning mechanisms and to require later formation of noun–verb semantic relationships.

An alternative possibility (the “Early Interaction hypothesis”) is that noun-noun, verb-verb, and noun-verb semantic relationships are present from the onset of vocabulary development. Some authors suggest that word meaning is inherently and promiscuously relational. For example, Elman (2009) argues that word meaning arises from the multilevel contexts in which words appear, whether those contexts be linguistic or situated in real-world events; in this way, word meanings cannot help but be connected semantically to one another, at the very least through simple associative or thematic connections that may be accessible to young toddlers with unsophisticated vocabulary knowledge. Other theories of word learning also emphasize that word learning happens in rich contexts and incorporates many sources of available information (Hoey, 2012, Hollich et al., 2000), raising that possibility that semantic linkages among words are built into their meanings from the beginning of word learning.

For noun–verb semantic relationships in particular, evidence consistent with this interactionist perspective comes from the fact that in adults, verb meaning is tied to the participants involved in associated events – participants’ foregrounded temporary or permanent semantic features (McRae et al., 1997). Semantic properties of participants in events are also quickly processed and stored (Griffin and Bock, 2000, Hafri et al., 2013, Hafri et al., 2018, Ünal et al., 2019). Adults know, for instance, that part of the meaning of the verb “eat” involves understanding what can (e.g., “cookie”) and likely cannot (e.g., “clouds”) be eaten. Given that words are usually learned from real-world events in toddlerhood, relationships among the participants in events may directly form part of children’s noun and verb knowledge from the onset of word learning; a child’s learning of the word “eat” from a cookie-eating event would immediately be supported by and shape their knowledge of the word “cookie” and other foods.

Further evidence for the Early Interaction hypothesis for noun–verb semantic linkages comes from work showing that children’s early verb productions often occur in collocations with other nouns (Theakston et al., 2012, Theakston et al., 2015; cf. Naigles et al., 2009), suggesting a close semantic link between specific verbs and specific nouns in early verb use. Other work, cited above, shows relationships between vocabulary size and the ability to engage in verb-based prediction of upcoming nouns in sentence processing (Borovsky et al., 2012, Fernald et al., 2008). The association between vocabulary growth and use of noun–verb semantic linkages in sentence processing is further supported by a longitudinal study of language development in 15–24-month-olds that shows that children’s online word recognition skill closely tracks their verb-based sentence prediction skill (Reuter et al., 2023). This latter study shows that word- and sentence-level language comprehension skills develop together, suggesting that noun and verb semantic knowledge may also impact each other from the beginning of word learning.

After noun and verb semantics begin to be linked within and between parts of speech, how might these links affect later noun and verb learning and processing? Studies of noun vocabulary learning and processing indicate that the level of analysis is an important consideration – that is, whether analyses consider local associations between single words, associations among neighborhoods or groups of words, or even the global similarity of a word to all the other words in a vocabulary. Patterns of semantic association at different levels can have different effects on word processing. For example, Borovsky (2022) found that toddlers’ lexical recognition was differently associated with measures of words’ semantic association at more local compared to more global levels. Similarly, Mirman and Magnuson (2008) found that adults’ visual word processing was slowed by high density of words in nearby semantic neighborhoods but facilitated by high density in more distant semantic neighborhoods.

Systematic investigation into how word semantic association at different levels affects word learning and processing has often been based on network science methods. Such methods provide a detailed look at the semantic interrelationships among the words in children’s vocabularies across development at a large scale, revealing that vocabularies’ semantic structure influences word learning and processing. This “semantic structure” encompasses such factors as the number of semantic characteristics locally shared between words (Engelthaler and Hills, 2017, Sailor, 2013), the interconnectedness of words within a semantic neighborhood (Borovsky et al., 2016a, Mirman and Magnuson, 2008), and larger-scale features like semantic distance to many other words (Peters and Borovsky, 2019, Stella et al., 2018). Beyond children’s overall vocabulary development as indexed by vocabulary size, aspects of semantic structure like large-scale network connectivity can also differentiate late talkers from typically developing children (Beckage et al., 2011, Borovsky et al., 2021), help in characterizing early vocabulary growth strategies (Hills et al., 2009), and modulate word processing in children (Borovsky, 2020, Peters et al., 2021). Accumulating evidence also suggests that the semantic structure of children’s vocabularies at different levels of analysis influences the order in which nouns are learned (Beckage and Colunga, 2019, Borovsky et al., 2016a, Engelthaler and Hills, 2017, Fourtassi et al., 2020, Hills et al., 2009, Hills et al., 2010, Peters and Borovsky, 2019, Sailor, 2013, Stella et al., 2018, Steyvers and Tenenbaum, 2005).

Almost all of the vocabulary modeling studies using tools from network science have focused on nouns, and those that have included verbs have analyzed overall vocabulary structure rather than the semantic interrelationships between nouns and verbs (e.g., Beckage et al., 2011, Dubossarsky et al., 2017). As such, it is not clear how the cross-level patterns of semantic structure described above may differently affect noun-noun, verb-verb, and noun-verb semantic interrelationships. One possibility is that the dynamics of semantic interconnectivity across levels of semantic structure reflect large-scale cognitive mechanisms that apply to all kinds of words, regardless of part of speech. This would imply that across levels of semantic structure, semantic relationships within and among nouns and verbs would exert similar effects on noun and verb processing and learning. Conversely, it is possible is that semantic connectivity at different levels would differently affect noun-noun, verb-verb, and noun-verb semantic interrelationships due to the inherent differences between noun and verb semantics.

In this study, we aimed to understand how semantic interrelationships within and between nouns and verbs across levels of semantic structure develop over time and influence subsequent noun and verb learning. We asked whether nouns and verbs would demonstrate patterns of semantic structure in early vocabulary development that were more consistent with the Early Isolation hypothesis or the Early Interaction hypothesis. We also assessed how these aspects of semantic structure, once they emerged, affected later word learning.

In our approach, we modelled children’s vocabularies as networks of words. In these networks, words are treated as nodes and shared semantic features serve to connect words through edges in a network (see, e.g., Steyvers & Tenenbaum, 2005). Semantic features are defined as semantic subcomponents of word meaning (McRae et al., 2005, Vinson and Vigliocco, 2008). For example, the semantic features associated with the word “dog” include < has fur >, < eats bones >, and < is a pet >. Semantic features have been used to describe language processing and learning in numerous studies of both adults (e.g., McRae et al., 1997, Pexman et al., 2008) and children (e.g., Hills et al., 2009, Peters and Borovsky, 2019, Stella et al., 2018). We examined the properties of the nouns and verbs in children’s individual vocabulary networks at three levels of semantic structure: the word, neighborhood, and lexicon levels (see Fig. 1 for a general illustration of the levels and Fig. 2 for examples of the associated network measures).

We contrasted noun-noun, verb-verb, and noun-verb semantic interrelationships by separately measuring nouns’ and verbs’ semantic structure associated with other nouns (i.e., their noun-specific semantic structure) and their semantic structure associated with other verbs (i.e., their verb-specific semantic structure). Note that nouns and verbs have both noun-specific structure (i.e., relationships with other nouns) and verb-specific structure (i.e., relationships with other verbs); measuring these different relationships allowed us to separately consider relationships within and across parts of speech. Because these relationships are considered from the perspective of individual words in a network, the relationships are not necessarily symmetric. For example, a verb’s noun-specific structure (e.g., eat’s connections with “bagel”, “cookie”, and “cake”) may look very different from a connected noun’s verb-specific structure (e.g., cake’s connections to “eat”, “slice”, and “bake”). Therefore, to assess noun-verb semantic linkages, we separately measured how nouns’ connections to other verbs may support noun learning and how verbs’ connections to other nouns may support verb learning. Finally, we assessed how these relationships might change over time by measuring semantic structure cross-sectionally in vocabularies of different sizes and longitudinally during vocabulary growth.

Below, we consider prior work bearing on whether noun and verb semantic structure and learning might differ across levels of semantic structure over time. We first consider how noun and verb semantic structure and learning might be affected at any time during vocabulary development by semantic connections to other nouns and verbs and then consider the evidence for the developmental timing of the emergence of these effects. (See Fig. 3 for an outline of the hypotheses.).

Word level structure encompasses the semantic connections shared between a word and its neighbors. For example, the word “dog” and “cat” share a number of semantic features, including < has fur > and < is a pet >, and are therefore semantically connected. The word “dog” also shares semantic features with the word “alligator”, including < is a carnivore > and < has a tail >. The network measure degree (k) quantifies the number of directly connected semantic neighbors to each word (Newman, 2018). In Fig. 1, degree for the red node (k = 4) includes relationships between the red node and its blue neighbors.

Word-level structure has been implicated in a number of studies of early noun learning (e.g., Peters and Borovsky, 2019, Sailor, 2013). For example, Peters and Borovsky (2019) found that toddlers tended to produce nouns with high degree at younger ages. These effects have been argued to occur because words with more connections or features are more semantically “rich” or “accessible” and facilitate spreading activation in the child’s lexical network (Peters and Borovsky, 2019, Pexman et al., 2008). Other work has emphasized that a noun’s distinctive features (i.e., those that it does not share with other words in the network) can also facilitate learning, suggesting that, for nouns at least, both connections to other nouns and semantic uniqueness may be important (Engelthaler and Hills, 2017, Siew, 2020).

Similarly, semantic connections to already-known verbs might also be helpful in noun learning. Support for this idea comes from work reviewed above that even very young children can use knowledge of the selectional restrictions of verbs to predict upcoming nouns (Fernald et al., 2008, Mani and Huettig, 2012, Reuter et al., 2023). Other work indicates that children’s learning of novel nouns can be supported by the overall semantic properties of verbs’ argument structure (Lidz et al., 2017) and also by the properties of individual verbs (White & Lidz, 2022). This latter study provides particularly strong support for the idea that learning a noun with semantic connections to already-known verbs is easier, potentially because children will be better prepared to identify its referent in the world from the sentences they hear.

Verb learning may also benefit from semantic connections to already-known nouns. For example, children are more likely to learn a verb when they are familiar with the participants (i.e., nouns) associated with the action which the verb describes (Kersten & Smith, 2002) and when those participants are explicitly named and semantically detailed (Arunachalam and Waxman, 2011, Arunachalam and Waxman, 2015, Gleitman et al., 2005). Other work shows that language input that includes sentences containing a diversity of nominal subjects – as opposed to few subjects or only pronominal subjects – is associated with increased diversity in the number of different verbs (and subjects) in child language production (Hadley et al., 2017). Robust noun knowledge and input may help children come to an event prepared to grapple with the task of learning a new verb without needing to account for unknown nouns as well.

However, it is not clear whether or how semantic connections to other verbs might support verb learning. Prior evidence suggests that verb learning might be either helped or hindered by strong semantic connections to other verbs. On one hand, some work has examined the possibility that certain early verbs serve as “pathbreaking” verbs by establishing an abstract argument structure pattern that is then used to facilitate the learning of later-learned verbs (Ninio, 1999; cf. Naigles et al., 2009). For example, a generic verb like do might help establish the subject-verb-object abstract pattern which is then used by other verbs. Relatedly, learning a novel verb might be facilitated by knowledge of semantically similar verbs (ones that might also have similar argument structure: e.g., x eats y, x drinks y; Levin, 1993). Additionally, verb vocabulary growth during toddlerhood is predicted by the diversity of verbs heard in the language input (Hsu et al., 2017). These findings suggest that verb learning may be aided by semantic connections to a variety of other verbs.

On the other hand, children can take a long time to learn the distinctions between different verbs within a semantic domain. For example, in Mandarin, where there are over a dozen verbs that can be translated into English as carry or hold, even seven-year-old children have some difficulty differentiating verbs within this crowded semantic domain (Saji et al., 2011). Further, these children tend to first use verbs with more general carrying or holding meanings, indicating that while the broad properties of this semantic domain are learned early, differentiation into more specific meanings takes more time. This finding suggests that learning a verb with many direct semantic connections with already-known verbs may be a difficult task for young children. This challenge may be compounded by the difficulty of associating verbs with actions from events, a difficulty experienced even by adults. For instance, when shown videos of toddlers and parents playing and talking together with either no audio or with the verbs in the parents’ speech masked by a beep, adult participants were significantly worse at guessing the verbs uttered by the parents compared to the nouns (Gillette et al., 1999); work using head-mounted cameras showing children’s perspectives during play has similar findings (Zhang et al., 2022). Together, these findings indicate that a child with a growing verb vocabulary may experience interference from other known verbs because the semantic boundaries between verbs may be imprecise.

Neighborhood level structure characterizes the relationships among groups of words such as members of a specific taxonomic category like animals or vehicles. In the current study, a network measure called local clustering coefficient is used to quantify relationships at the neighborhood level for an individual word. Local clustering coefficient (hereafter called clustering coefficient for brevity) measures the density of interconnections among a word’s neighbors. In contrast with degree, which measures a word’s direct connections to other words in the network, clustering coefficient measures the propensity for a word’s neighbors to also share connections themselves. This is often calculated as a proportion, in which the denominator is the number of possible connections that could exist and the numerator is the actual number of connections that do exist. In Fig. 1, this measure captures the structural relationships among the neighbors (in blue) of the word of interest (in red) and the nodes that the neighbors connect to (in green).

For nouns, several studies suggest that noun-related neighborhood level structure can support noun learning and processing. For instance, at 24 months, children who produce more nouns in a superordinate category show facilitation when learning a novel member of that category (Borovsky et al., 2016a, Borovsky, 2020). Though relatively less explored, semantic density appears to facilitate noun recognition in very young children (who have smaller vocabulary networks than adults). For example, category-level density in noun neighborhoods facilitates noun recognition in children between 18 and 24 months of age (Borovsky, 2022, Borovsky et al., 2016b, Borovsky and Peters, 2019). How noun learning interacts with prior verb neighborhood structure is not clear, though it is possible that noun learning would be aided by verb neighborhood structure just as it is by noun neighborhood structure.

Verb learning may also be facilitated when verbs connect into well-developed neighborhoods of nouns. Evidence for this hypothesis comes from research showing that children demonstrate better extension of verbs learned with perceptually and semantically similar objects than with dissimilar or single objects (Childers et al., 2016, Haryu et al., 2011, Kueser et al., 2023). Kueser et al. (2023) also showed that the ability to extend verbs learned from events involving objects from a single semantic category was positively associated with children’s knowledge of that semantic category. This suggests that highly interconnected noun neighborhoods may facilitate verb learning.

With respect to verb-verb semantic relationships at the neighborhood level, however, our expectations are less clear. If children have difficulty distinguishing between verbs with similar meanings (Saji et al., 2011) or have difficulty identifying the verb referent from events (Gillette et al., 1999, Zhang et al., 2022), then learning a verb that connects into highly interconnected verb neighborhoods may be a difficult task. Nevertheless, if a child already possesses a vocabulary with a highly interconnected verb neighborhood, that child might be ready for the task of learning another verb with a similar meaning. Children benefit from implicit and explicit contrasts among novel events and verbs (Childers et al., 2014, Imai and Childers, 2020), and densely structured verb neighborhoods may support this type of contrastive learning.

Measures of a network at the lexicon level consider a word’s position with respect to all of the other words in the vocabulary. This study uses a network measure called betweenness centrality, which measures whether a word is along the shortest paths between other words. For example, in a network consisting of the words “hay”, “eat”, and “horse”, “eat” would have high betweenness centrality because it would bridge the two otherwise unconnected words “hay” and “horse”. In Fig. 1, measurements of the red node at the lexicon level include its relationship to all of the nodes in the network (in blue, green, and yellow).

Work suggests that high noun-specific (and potentially verb-specific) betweenness centrality may help noun learning. For example, betweenness centrality predicts an individual noun’s entry into children’s vocabularies (Beckage & Colunga, 2019) and ranks highly as a predictor of later language outcomes (Borovsky et al., 2021).

Verb learning in particular might benefit from high levels of betweenness centrality with respect to other nouns, given that verbs serve a syntactically “bridging” role in sentences. For example, if a child already knows many of the agents and patients associated with the subject and object positions of a transitive verb’s argument structure, then that verb would likely occupy a central, hub-like position in the lexicon as a whole. Learning such a verb may not be particularly difficult because when hearing sentences containing that verb, extracting the meaning of the verb would be among the child’s only tasks. This situation would contrast with a transitive verb where, say, only one of the verb’s arguments is known; in hearing a sentence with such a verb, children would need to extract the meaning of the verb and the other argument, a more difficult task. In fact, studies of verb vocabulary acquisition over toddlerhood have suggested that verbs with more arguments tend to be learned later(Horvath et al., 2021, Horvath et al., 2018, Horvath et al., 2019). If verbs need high noun-specific betweenness centrality to be learned, then verbs with more arguments will require more developed semantic structure in multiple noun neighborhoods, delaying their acquisition compared to verbs with fewer arguments. The impact of a verb’s betweenness centrality with respect to other verbs is likely less important given that the syntactic structure of the English language does not tend to create utterances in which one verb is surrounded by other verbs.

Semantic structure effects across levels may differ in their emergence across development. For example, we may find evidence consistent with the Early Isolation or Early Interaction hypothesis at one level and not at other levels. Intuitively, we might expect the most support for the Early Isolation hypothesis at the neighborhood and lexicon levels – as larger-scale semantic interconnectivity may take more time to develop – and the most support for the Early Interaction hypothesis from the patterns of local semantic connections seen at the word level. Timing differences may also be observed in how early semantic structure across levels affects subsequent word learning.

The current study assesses how nouns and verbs demonstrate semantic structure within and between parts of speech, focusing on noun-noun, verb-verb, and noun-verb relationships. We examine two competing hypotheses regarding the development of semantic structure: the Early Isolation hypothesis, which claims that early semantic representations are isolated from one another, and the Early Interaction hypothesis, which claims that early semantic representations interact. We look for evidence of semantic interaction across three levels of analysis – the word, neighborhood, and lexicon levels. We performed two experiments that offer complementary views of the development of semantic structure to answer two questions:

• 1. Do nouns and verbs display large-scale cross-sectional semantic relationships across levels of semantic structure in early vocabulary development consistent with the Early Isolation hypothesis or the Early Interaction hypothesis?

• 2. Do nouns’ and verbs’ early semantic relationships affect later noun and verb learning in ways consistent with the Early Isolation hypothesis or the Early Interaction hypothesis?

In Experiment 1, we answered the first question using a cross-sectional analysis to measure vocabulary structure related to nouns and verbs at the word, neighborhood, and lexicon levels in 16–30-month-old toddlers with a wide range of vocabulary compositions. We used cross-sectional parent checklist data describing these children’s early vocabularies and created semantic networks for each child. With measures of vocabulary structure for each word in the child’s vocabulary, we assessed three characteristics of children’s vocabularies. First, we assessed whether children’s vocabularies displayed evidence of a systematic response to semantic structure across levels, that is, whether the structure observed in actual children’s vocabularies differed from that seen in synthetic vocabularies based on randomly generated networks with the same numbers of nouns and verbs to control for the larger numbers of nouns in early vocabularies (Frank et al., 2021). Second, we measured how individual nouns and verbs were structurally positioned with respect to noun- compared to verb-specific networks in the children’s vocabularies. Third, we measured the extent to which the structural positions of nouns compared to verbs differed within the noun- and verb-specific networks. Together, these analyses addressed our first question about how vocabulary structure with respect to nouns and verbs changes over development. Support for the Early Isolation hypothesis would be found in patterns in which words systematically demonstrate no more or less semantic connectivity than expected compared to random networks. In contrast, support for the Early Interaction hypothesis would be found if nouns and verbs demonstrated systematically more or less semantic connectivity than expected from the onset of vocabulary development. To preview the results, we found the most evidence for the Early Interaction hypothesis and found varied patterns of semantic interconnectivity within and between nouns and verbs.

To confirm whether and how early semantic interactions among nouns and verbs shaped children’s later vocabulary development, we performed a longitudinal analysis in Experiment 2 to address our second question about how early noun and verb vocabulary structure affects later noun and verb learning. In contrast with the cross-sectional analysis in Experiment 1, in Experiment 2, we used a longitudinal design that supported inferences about the directionality of the effect of network structure on subsequent vocabulary growth. We followed an approach used in prior work with nouns in which a longitudinal vocabulary network was grown by adding words to the network in the order in which they were learned across a sample of children (i.e., their normative order or age of acquisition; Hills et al., 2009, Peters and Borovsky, 2019). We then measured how earlier noun and verb vocabulary structure contributed to subsequent noun and verb learning.