Chapter 17 Language Development at Scale
This chapter synthesizes knowledge gained from the broader enterprise, attempting to identify generalizations about early language development that hold true across the different datasets in our sample. We begin by briefly reviewing the content of our findings across the preceding substantive chapters. Next, we discuss three synthetic generalizations from these findings. We then turn to the question of particular process universals that might underly this empirical picture.
We began with the question of what parent report can tell us about children’s early language. We are now in a better position to answer this question, summarizing findings across the content chapters of the manuscript.
In Chapter 4, we reviewed evidence for the reliability and validity of parent report instruments for assessing children’s early language, with generally positive conclusions. Further, we contributed two sets of novel analyses. First, examining large-scale longitudinal datasets, we estimated the fall-off in test-retest correlations over developmental time, which suggested a relatively high level of reliability overall. Second, we used item response theory (IRT) models to examine the measurement properties of individual words on the CDI. Overall, most items had strong psychometric performance, and even comprehension reports appeared to have consistent informational value about a child’s overall abilities. The broad picture from this chapter was positive with respect to the psychometric properties of the CDI, motivating further use of these data.
Chapter 5 then examined trends in the growth and distribution of individual vocabulary estimates. Across languages, and for production and comprehension, early vocabulary development starts slow and grows rapidly in the second year after birth. There are both absolute and relative differences between languages in the estimated rate of vocabulary growth, however. While some of these may reflect true differences (for example, Danish being famously difficult to acquire, at least at the early stages; Bleses et al. 2008), others likely stem from differences in instrument, sample, and administration conditions. Cross-language variability was dwarfed by cross-individual variability. Further, we observed a striking consistency in this variability across languages. On average, language emerges quickly – but individual toddlers around the world are highly variable in their own learning rate.
Chapter 6 considered demographic predictors of vocabulary size. We found a general, though small, female advantage in early vocabulary that was more pronounced for production than comprehension. This advantage did not appear to stem from reporting bias, and was relatively consistent in size across languages. We also observed a first-born advantage and an advantage for children with more-educated mothers (treating maternal education as a proxy for general socioeconomic status). Chapter 9 followed up these analyses by identifying whether specific words were the locus of some of these differences. The conclusion of these analyses were that there were both general demographic differences in vocabulary and word-specific differences. For example, girls, on average, both have a higher probability of knowing any word, and they also have a much higher probability of knowing words like dress and doll.
Turning to gestural communication, Chapter 7 showed that children’s uses of early gestures are both quite consistent with one another and quite consistent across languages. Gestures appear to be acquired together as part of a suite of communicative abilities, rather than being learned piecemeal as isolated items: A small set of similar gestures are often the earliest emerging across languages, likely arising from relatively universal dyadic needs (e.g., a signal for “pick me up” or “give me”). Finally, within individuals, variation in early gestures is highly correlated with variation in early vocabulary.
Like early gestures, early words are often surprisingly similar across languages. In Chapter 8, we quantified this similarity and showed that it was related to typological relatedness. On the other hand, a set of core concepts were encoded by words in children’s earliest vocabulary, even across quite dissimilar languages. Like early gestures, these early vocabulary items likely reflect relatively universal communicative needs in infant-caregiver dyads. (We return to this theme below in Section 17.2).
In Chapter 10, we created predictive models of the ease or difficulty of individual words. These models take into account the average environmental input for a learner of a particular language (estimated from corpus resources). Although a number of factors including input, conceptual factors, and phonological complexity interact to predict when words are first understood or produced, the profile of predictors across languages was quite similar. This finding points to a relatively consistent set of processes that operate for children learning different languages in different contexts.
Following the thread of cross-linguistic differences, Chapter 11 examined the question of the syntactic composition of vocabulary across early development. Consistent with previous reports, we found that most languages exhibit a bias towards nouns, although the strength of that bias varied across languages. Further, essentially all languages showed a developmental bias against function words. On the other hand, the bias for predicates (verbs and adjectives) was more variable across languages, with most languages showing a negative bias, but a few (Mandarin and Cantonese, in particular) showing a neutral or positive bias. This pattern of variability points to some language-specific factors – perhaps syntactic structure, perhaps other linguistic or interactional factors – that influence learning of words from particular syntactic categories.
We next applied this same approach to semantic categories (Chapter 12). Computing the relative bias for or against particular semantic categories, we found some cross-linguistic consistency in early biases: biases for body parts, games, and onomatopoeia were quite consistent; biases against words for places and time were as well. These biases suggest that there are likely some attentional and conceptual biases that facilitate or inhibit word learning in similar ways across languages.
Turning next to syntactic development, Chapter 13 showed that children’s morphological and grammatical ability appears tightly coupled with their vocabulary size. This generalization holds very consistently across languages, providing support for Bates’ speculation that there is a general core of language development that is shared across vocabulary and grammar learning. This chapter extended earlier analyses of these relations by identifying a moderating effect of age on this relationship such that older children appear to gain more “units grammar” per “unit vocabulary” – that is, grammar emerges slightly faster for these older, compared to younger, children. Further, analyses of longitudinal data suggested that, for children in the CDI age range, vocabulary learning is temporally prior (and perhaps, causally prior as well) to grammatical complexity, supporting a view in which grammatical abilities emerge from generalizations based on learned vocabulary.
Chapter 14 examined children’s morphological overgeneralizations specifically. Here, the data supported the idea of a phase of over-regularization and recovery, with some evidence that over-regularization was correlated within children across distinct morphological generalizations in nouns and in verbs.
The last empirical chapter, Chapter 15, examined individual variability in the ways children approach the task of learning or “style.” This chapter presented evidence for the “referential” vs. “expressive” distinction in children’s learning trajectories across languages. Even controlling for overall developmental level, some children appear to have larger, “noun-ier” vocabularies, while others have smaller vocabularies, but are more likely to be reported to combine words. Similarly, some children certainly produce relatively more of the words they understand than other children, though these measurements are difficult to validate independent of reporting factors. On the other hand, we did not observe strong evidence for the idea that children’s vocabulary growth is particularly “spurt-y” – rather, once statistical artifacts are accounted for, vocabulary growth appears to accelerate relatively smoothly in all children, albeit with variation in acceleration rates across children.
What is the picture of language development that emerges from these individual findings? Making use of the patterns of cross-linguistic variability and consistency that emerged in Chapter 16, we now return to the theoretical project of Chapter 1, which was to use these patterns to provide constraints on theories of early language learning. We begin by considering three major generalizations from our data, and then turn back to the question of possible learning processes that could support these generalizations.
17.2.1 The language system is tightly woven
It could easily be the case that children are “pointy” with respect to language – that is, highly proficient in some parts and less proficient in others. A visitor to a toddler classroom might plausibly observe the following pattern: some children primarily gesturing to communicate; some others knowing the names of many things but not yet combining words; and still others fluently combining words into multi-word utterances. Furthermore, these children could be distributed across a range of ages, with language ability only somewhat predicted by relative age. Seeing these differences, this visitor might conclude that the variation they observed was supportive of a multi-factorial viewpoint, where some children are better at words, others are better at grammar, and still others are better at gestures.
While there is some true variation of this sort (see point three below), the kind of observations described above will tend to overstate it. Instead, our analyses suggest that the language system is tightly woven together, such that, on the whole, children who are good at gestures tend to have larger vocabularies, and children who have larger vocabularies also tend to inflect and combine words more proficiently. These correlations within individuals are large and they are almost completely consistent across the languages in our sample. Further, they are also largely borne out in a variety of observational and behavioral datasets that do not rely on parent report (e.g., Brinchmann, Braeken, and Lyster 2018; Rowe and Goldin-Meadow 2009). How can the consistency and coherence of these distinct aspects of early language be reconciled with the type of variation a preschool observer might see?
Consider a model in which learning follows a single, coherent path through a set of discrete stages. Initially, gestures and early nouns and social routines are learned. Nouns follow, leading into verbs. Verbs in turn promote the use of word combinations and morphology. In such a model, children’s rate of learning those particular features of language could be non-linear. For example, their noun bias could increase and then decrease just by virtue of whether they had begun learning verbs (Bates et al. 1994). And, the relation between lexicon and grammar could be flatter at one period than another (Dixon and Marchman 2007).
Under this kind of model, the children in the preschool classroom could simply be at different points along the same trajectory. A younger child who knows a surprising number of names might be further along this trajectory relative to her age and hence firmly in the “noun bias” part of the general path. In contrast, an older child who primarily gestures might be a child with a slightly slower growth rate. In other words, observed variation need not indicate that the sub-parts of language do not “hang together” (to return to the Batesian phrasing).
Alongside its support from the cross-linguistic consistency in cross-domain correlations, this unitary view of early language is supported by several other bodies of work. The first is a set of longitudinal analyses by Bornstein and colleagues. In a two-cohort longitudinal study, Bornstein and Haynes (1998) and Bornstein, Hahn, and Haynes (2004) collected data from a sample of American English learning children at two and four years and measured language using a variety of instruments including parent report, transcripts, and behavioral tasks. In a re-analysis of these data, Bornstein and Putnick (2012) found that the core construct of early language that emerged from the sub-measures was both highly correlated with each of the sub measures and also highly stable over time.
Research on international adoption provides additional support for this unitary view. International adoptees often have to let go of their native language completely and begin learning a host language, sometimes years after they have begun acquiring their native language. Work by Snedeker, Geren, and Shafto (2007, 2012) suggests that these children pass through the same general stages of acquisition as learners with a single native language, but that they do so much more quickly on average than native learners.
In sum, the language system is much more tightly woven than casual observation might lead an observer to expect. Although our data provided multiple opportunities for aspects of language development (communication, vocabulary, morphology, grammar) to dissociate into modular subsystems across individuals, measures of ability instead hung together very tightly. Further, the strength of these connections varied little across the languages we examined.
17.2.2 Children’s similar interests drive their communication
A second broad generalization from our work here is the content similarities in children’s early communication – both vocabulary and gesture – across languages. Put simply, regardless of the language they are learning, children in the early stages of language learning appear to talk about similar things, especially people, social routines, small objects, and body parts. This work builds on analyses by Tardif et al. (2008), who compared the distribution of children’s first ten words across three languages, Mandarin, Cantonese and English. Many of the content generalizations that held for those three languages in fact hold here as well in a much more broad array of languages. Further, children’s early gestures are also strikingly similar across languages, communicating showing, giving, and the desire to be picked up before their first birthday with high consistency across languages.
A naive observer might suppose that children in different cultures begin communicating about quite different things. They could imagine a culture in which children primarily begin by describing the properties of objects (“soft!” “small!”) or another in which children first describe places or things around the house. Or, perhaps it is just American children who are obsessed with animals, while children in other cultures are less so. But, it turns out that children in different cultures are much more similar in the content of their early communication than we might have expected.33
We also observed a strong correlation in the order of acquisition for the earliest words and gestures. Those words and gestures that were typically learned very early in one language tended to be very early learned in others. Correlations in the ordering of words were attenuated for later-learned words, however. Looking at the vocabulary as a whole, we also saw that some semantic categories were over-represented across languages, including, words for body parts, games and social routines, and sounds. (In contrast, predicates and function words, as well as words for time and places were under-represented.) When we looked for systematic predictors of the ease of learning words across language, the predictor “babiness” – which captures association with infants – was a strong predictor, especially for early-learned words.
Where do these similarities come from? They likely do not emerge from simple linguistic or environmental frequency. Not only is frequency statistically controlled in some of our analyses (e.g., those in Chapter 10), but also there are clear dissociations between linguistic frequency and children’s comprehension and production. The words for mother and father are relatively infrequent in language directed to children – yet these words are learned very early. In contrast, many function words are highly frequent, yet are produced much later than expected. And many omnipresent environmental stimuli from couches and carpets to diapers are referred to only much later than the people, animals, and small objects on and in them.
Instead, similarities in children’s early communications across languages likely emerge from a combination of factors: (a) children’s particular attention and interests; (b) the communicative priorities of child-caregiver dyads; and (c) the informational structure of language. These factors work together to render certain aspects of language easier to learn than others.
What children are interested in and what they pay attention to are powerful determinants of children’s behavior, as anyone who has spent time with toddlers will likely attest. The red puppet Elmo, who has a squeaky (some might say annoying) voice, is optimized to draw children’s attention and interest, despite some parents’ best efforts to minimize his frequency in the environment. More generally, toddlers tend to make a bee-line for animals and toys that they perceive as interesting. These preferences may stem from preferences for animates, perceptual preferences (e.g., for bright red and high voices), or a mixture of these, but their impact on children’s attention – and from there, their vocabulary across languages – is undeniable.
Although toddlers have their own interests, their language exposure takes place in an environment that is largely constructed by their caregivers. Much of the language they hear is “functional talk” that takes place in the context of routines like dressing, diapering, mealtime, and bathtime. These dyadic priorities mean that children’s vocabulary often contains words that help them name and navigate these routines. Roy et al. (2015) even argued that the contextual distinctiveness of words that appear in these routines – signaled by their distribution in space, time, and linguistic context – may lead to earlier acquisition. Across cultures, if parents are going through the same functional routines, this similarity would lead to cross-linguistic consistency in what words their children learn.
Finally, some words may be learned earlier than others simply because of the informational structure of language (Snedeker, Geren, and Shafto 2007). Verbs may be learned later than nouns simply because you need to know some nouns to figure out the meanings of the verbs (Gleitman 1990; Gillette et al. 1999). To the extent that these regularities extend across languages, they would impose a variety of ordering constraints on acquisition that could be reflected in our analyses.
17.2.3 Children take different routes into language
In the original CDI norming study monograph, Fenson et al. (1994) noted that variability is perhaps the primary and most striking fact about children’s vocabulary learning. Our observations confirm this conclusion: Despite the aggregate similarities across children and across cultures, the speed with which children acquire language is highly variable in the first years of life. From a biological perspective, this variability is quite unprecedented. As a comparison, variation in heights for toddlers is tiny compared with variation in vocabulary: the mean height for a 24-month-old is around 33 inches, with a standard deviation of a little more than an inch, leading to a coefficient of variation around .03. This measurement is almost two orders of magnitude smaller than the coefficient of variation on vocabulary.
Where does this variability come from? We can only speculate. Some must come from variation in input across households, which has been amply documented to relate to children’s early vocabulary (e.g., Hart and Risley 1995; Hoff 2003; Weisleder and Fernald 2013). Further, some component of this input-correlated variation is likely to be genetic (e.g., Hayiou-Thomas, Dale, and Plomin 2012), such that some children inherit a tendency towards slower vocabulary growth from parents who themselves talk relatively less and use relatively less diverse vocabulary (Dale et al. 2015).
The degree of variability we observed itself is a constant across cultures, however. Examining the variability in English-learning children’s vocabulary documented by Fenson et al. (1994), it was easy to think that the spread of children’s outcomes was due to the demographic and parenting variability found in the United States, which – even in the restricted set sampled in that initial study – was large. But, a look at the variability estimates found in our broader sample quickly falsified this hypothesis.
Further, children vary stably in the nature of their acquisition path, some naming more objects and others combining words relatively more. Controlling for the non-linear trajectory of acquisition described above, we found stable differences in “referential style” among children. Although the language system is tightly woven and moves through a relatively consistent learning trajectory across individuals, there is nevertheless an interesting, second-order component. Some children, especially the faster-learning ones, appear to learn more nouns. Others, often the slower-learning ones, tend to combine words more frequently. As we saw in Chapter 13, it is these older children who appear to gain more “grammar per unit lexicon.” Maybe these early word combinations are actually “unanalyzed wholes” that are syntactically less complex than they might appear. Perhaps those children who are learning more slowly are then able to bring more mature working memory to the task of grammatical induction. Or perhaps these children have heard more language and hence the natural statistics of language are more apparent to them. Regardless, these differences appear as a stable aspect of the relatively consistent developmental course we observed.
Does variation in rate or style of learning persist beyond the range of our study? One of the most intriguing aspects of our rate analysis was the suggestion that, accounting for ceiling effects, variation in ability does not compress in the range we measured. An important direction for future work would be to ask about the range of variation observed within and across cultures on other standardized measures of language in school-age children. One possibility is that less variability will be observed for older children on most standard vocabulary and grammar measures simply because functional communication is possible for nearly all speakers – ceiling effects, in essence. Instead, variation will be measurable in more “leading edge” domains like literacy or discourse comprehension.
17.3 Learning processes
In Chapter 1, we introduced the idea of “process universals.” These cannot be universals of content as all of the content being reported by parents filling out CDI forms is language-specific. Instead, similar to Slobin (1973)’s “operating principles,” we are interested in processes that operate in different language contexts to produce the observed pattern of phenomena (cf. Clark 1977). What more can we say about potential process universals, building on the generalizations above? We highlight three process-level connections that appear consistent with our data.
17.3.1 Language grows through interactional input
Without input, there can be no uptake. The importance of language input is a fundamental tenet of all models of word learning, from the simplest accumulator model (McMurray 2007) through to more complex models (e.g., McMurray, Horst, and Samuelson 2012; Frank, Goodman, and Tenenbaum 2009; Fazly, Alishahi, and Stevenson 2010). In all of these models, what is learned by the model is a function of the frequency and statistical distinctiveness of the learner’s input. This conclusion is amply supported by a large body of correlational research linking observed speech from children’s caregivers to their vocabulary size (e.g., Hart and Risley 1995; Hoff 2003; Huttenlocher et al. 1991; Hurtado, Marchman, and Fernald 2008).
But, quantity is not enough. Beyond the first order correlation of input quantity to language outcomes, a body of research now provides nuanced qualification. In a number of studies across cultures, child-directed speech is a better predictor than overall speech, even in cultures where this kind of speech is relatively rare (Weisleder and Fernald 2013; Shneidman and Goldin-Meadow 2012). One hypothesis is that child-directed speech provides more grounded moments in which word meanings can be inferred from context (Cartmill et al. 2013). Indeed, in one study, a variety of measures of input quality – including joint engagement as well as the presence of rituals and routines – were better predictors of vocabulary size than pure quantity (Hirsh-Pasek et al. 2015). Of course, high-quality input must be developmentally appropriate, and for older children, language that is more syntactically complex supports complex syntax acquisition (Huttenlocher et al. 2002). Presumably the evidence for the importance of grounded, engaged communication is at least in part due to its specific importance for younger learners who are engaged in discovering word meanings (e.g., Clark 2007; Clark and Estigarribia 2011); other aspects of language gain importance for other learners (Hoff 2006; Hoff and Naigles 2002; Meredith L Rowe 2012).
We see many of our conclusions as fitting well with this broader picture. While we do not have predictors of individual children’s input, the research presented in Chapter 10 suggests that even aggregated, average measures of input are relatively powerful predictors of word-by-word uptake. The most obvious example of this is the repeated finding that word frequency is a useful predictor of age of acquisition, especially for nouns. Even if high-quality, grounded instances are the appropriate learning input, greater frequency overall will – all else being equal – lead to greater frequency in the appropriate learning context. We further observed effects of solo frequency (being used in a one-word sentence) and mean length of utterance. These results are both consistent with the idea that it is easier to learn meanings for words in shorter sentences, which pose both fewer word segmentation challenges and fewer word-meaning mapping ambiguities. In addition, the relative cross-linguistic consistency in these predictors suggests that the input-uptake connections that have largely been documented for English learners are likely to be robust for learners of other languages as well (though the strength this conclusion is moderated by the relative lack of typological diversity in the sample of languages we have available).
The demographic differences in vocabulary we observed are also consistent with interactional-input theories of vocabulary development, in which the more higher-quality the input the child receives, the faster vocabulary grows. Under this hypothesis, children who are first-born and who have mothers with more education are likely to receive more and more higher-quality input. First-born children receive more input through their greater allocation of parent attention (though as we note in Chapter 6, first-born children’s parents may also be more aware of their vocabulary). Children with mothers who have more education receive more and more higher-quality talk through the availability of more parent time, different values around talk, greater awareness of the role of interaction for young children, and perhaps differing parental practices (Evans 2004; Farkas and Beron 2004; M. L. Rowe, Suskind, and Hoff 2012).
17.3.2 Individual word meanings must be inferred based on (cross-situational) evidence
Cross-situational learning is the proposal that children use the statistical properties of how words are used across contexts to help them infer meaning (Gleitman 1990; Siskind 1996). This specific proposal has been instantiated in a variety of associative learning experiments with both adults (Yu and Smith 2007; Yurovsky and Frank 2015) and children (Smith and Yu 2008; Vlach and Johnson 2013). The specific mechanisms underlying learners’ performance in these tasks are still controversial (e.g., Medina et al. 2011; Trueswell et al. 2013; Yurovsky, Smith, and Yu 2013; Yurovsky and Frank 2015). Despite this controversy, the general model of learning as proposed in these studies has rapidly become the default instantiation of how interactional input translates into learning (e.g., the models of Frank, Goodman, and Tenenbaum 2009; Fazly, Alishahi, and Stevenson 2010; McMurray, Horst, and Samuelson 2012). The basic idea is that children bring multiple information sources – including statistical, social, and grammatical information – to bear on resolving the meaning of each utterance. Then, this information is accumulated and brought to bear on resolving reference in subsequent utterances (Frank, Goodman, and Tenenbaum 2009; Bohn and Frank 2020).
Several findings from our investigations are consistent with these general ideas. As noted above, the predictors of word difficulty we found in Chapter 10 are consistent with the general input-uptake viewpoint. In fact, it is very easy to see some of these effects being generated by learning in models from the general cross-situational learning proposal. Frequency effects are ubiquitous in cross-situational learning models (though their specifics vary somewhat from model to model; e.g., Kachergis, Yu, and Shiffrin 2012). Further, on the cross-situational perspective, shorter sentences and especially single-word utterances are less confounded in terms of the information they provide about the relations between words and their meanings, simply because they include fewer extra words that would create spurious co-occurrences with the referent of the sentence.
Above, we also speculated about the effects of children’s idiosyncratic interests and attention on their early communication. Such effects have been accommodated into models of cross-situational word learning, where attention and salience are posited to shape the mappings that are learned (Kachergis, Yu, and Shiffrin 2012; Roy and Pentland 2002; C. Yu and Ballard 2007). More generally, the viewpoint on cross-situational communicative mapping that we have advocated provides a convenient theoretical tool for integrating external, situational constraints on input with the children’s internal priorities. While external circumstances (e.g., caregivers’ priorities) provide the distribution of language that is heard, internal factors like motivation and attention can nevertheless shape what is learned from these distributions.
Further, the noun and verb bias findings reported in Chapter 11 can be accommodated in the cross-situational viewpoint as well. On this view, nouns are easy to learn because the more they are heard, the more opportunities children get to build consistent mappings between the words and their contextual referents. And, this regularity should be especially true for concrete nouns that are more likely to be found in grounded contexts (Gentner and Boroditsky 2001). In contrast, verbs and other predicates cannot be acquired as easily from basic co-occurrence. Syntactically “light” verbs like make or do require some nominal information to constrain their meaning in context. And “heavier” verbs may still be systematically ambiguous without syntactic information (e.g., chase/flee; Gleitman 1990). Thus, verb learning – and likely the learning of other predicates and many function words as well – relies on a base of nouns and a basic comprehension of the syntactic structures in which they appear in order to infer meaning in context (Gleitman 1990; Gillette et al. 1999). These effects appear in a probabilistic model of cross-situational noun/verb learning (Abend et al. 2017).
The information-sequencing viewpoint predicts that verbs should be acquired relatively later, with relatively more support from shorter, easier-to-parse utterances. Our finding that mean utterance length is a stronger predictor of acquisition ordering for predicates, rather than nouns, is consistent with this idea. This account explains cross-linguistic exceptions like Mandarin as cases where many early-learned verbs are semantically-transparent enough to be learned cross-situationally without syntactic information (following Tardif 1996). Though there is much more work to do here, the cross-situational viewpoint is a useful tool for exploring noun and verb biases in vocabulary development.
17.3.3 Generalizations appear gradually
Syntactic and morphological structures on the CDI are likely to be “item-based” in the sense of Tomasello (2003). Since all syntactic information is instantiated in particular example items, the knowledge that the CDI assesses is not as abstract as that posited by high-level syntactic theories. Although there may well be broader, more abstract generalizations that underlie the growth of word order, we simply do not have the signal to address the presence or absence of such abstractions. That said, our evidence is broadly consistent with the view of language as growing through the gradual induction of syntactic regularities through a reciprocal interaction with the acquisition of individual content words.
Versions of this viewpoints exist throughout the broad theoretical space of language acquisition (e.g., Yang 2016; Tomasello 2003; Meylan et al. 2017), but all proposals rely on a learning mechanism in which generalizations about structure are graded and rely on the amount of evidence available. All such mechanisms would presumably predict some relation between individuals’ grammatical and lexical abilities. An important target for future theoretical work in this area is to explore how tight these correlations are predicted to be on different viewpoints. Our suspicion is that only the most construction-based views will predict the level of coupling we observed, however (Bates et al. 1994).
The relation between grammar and the lexicon is reciprocal. As content vocabulary grows, it provides both the groundwork for generalization of constructions (Tomasello 2003; Goldberg 2006) and also the specific content words necessary for the induction of predicate meanings (as described above in the case of verbs). In the other direction, as grammar grows, it also enables better disambiguation of predicate meanings (Gleitman 1990) and creates myriad other opportunities for learning. Both of these directions of longitudinal influence can be observed, but in our study we found differential support for vocabulary supporting later grammatical ability. In contrast, Brinchmann, Braeken, and Lyster (2018) found support for grammar enabling vocabulary learning, although that study focused on older children. One plausible synthesis is that early grammar is based on generalizations from the growing vocabulary, while later vocabulary is acquired based on contextual inferences supported by grammar.
As we discussed in Chapter 1, we do not see our work here as resolving long-standing debates about the nature of abstract syntactic representations. Instead, we hope our contribution is somewhat different – we sought to refocus the debate away from phenomena that make only occasional contact with the gross regularities of children’s early language use in context. Instead, we focus on the pattern of observable changes in complexity and diversity of early language. We hope that this focus leads to theorizing – ideally accompanied by quantitative modeling – that takes these observables as its primary predictive target.
Developmental psychology often appears to be divided between two groups that do not communicate with one another. On the one hand, there are researchers interested in the ontogenetic and phylogenetic origins of knowledge – the epistemological project of understanding how we reason about objects, communicate using language, or learn from other people (Carey 2009; Spelke and Kinzler 2007; Tomasello 2010). On the other hand, there are researchers interested in growth, change, and variation across individuals in constructs like executive function, working memory, and personality (Diamond 2013; Gathercole et al. 1994; Ainsworth et al. 2015). The first type of research has led to a productive union of philosophical and psychological ideas, addressing exciting questions in the history of philosophy using empirical methods (e.g., Gopnik et al. 2004; Carey 2009). In contrast, the second type has had its impact in connections to clinical practice and to educational and public policy (e.g., Nelson 2007; Diamond and Lee 2011).
“Origins” research and “variation” research traditions have often appeared to be at odds with one another. While there are, of course, exceptions to this split, in many cases, researchers in these two traditions read different journals, record different measures, use different research designs and statistical models, and often appear to be pursuing different goals. Yet language learning is an area where these traditions come together. The knowledge being acquired – the stock of words in the child’s lexicon – is both the epistemic construct whose origins we are studying and the psychometric construct whose reliability across individuals we wish to assess. By studying the variability and consistency in the early lexicon, we can both study the origins of knowledge and the processes and factors by which human beings differ from one another across different cultural contexts, family environments, and genetic endowments.
In our work here, we sought generalizations in the universals and variations that hold across cultures and languages – and took steps to tie these generalizations to processes of language learning. These generalizations are based in the “variations” approach via the Batesian project of using variability to constrain theory. Yet, they are also about the “origins” of a very specific type of knowledge: knowledge of language. The ultimate aim of theorizing in this area must be broader than either “origins” or “variations” alone.
We hope that our work in compiling measurements of consistency and variability across languages is a beginning rather than an ending. Although the number of children represented in our analyses is large, the number of languages, and their diversity across families, is in the end quite small. And our models of acquisition – especially those linking input to uptake – are only a rough sketch of what is possible. Our hope, however, is that by showing what is possible in the quantitative study of language development, we illustrate a broader set of possibilities for a data-driven science of developmental change.
The caveats expressed in Chapter 1 still hold, however. Most of our languages are WEIRD (Henrich, Heine, and Norenzayan 2010), and the majority are Indo-European. A more diverse set of languages will be necessary for assessing the generality of our claims about the content of children’s early communication.↩︎