1. CHILDES AND THE DEVELOPMENT OF INTERROGATIVES

Among the modalities of the simple sentence that Brown (Reference Brown1973) identified as characteristic of his Stage III of development were interrogatives. An early article that focused on the development of one type of these, wh-questions, was Labov and Labov (Reference Labov, Labov, Campbell and Smith1978). This account of their daughter Jessie's learning of these forms over an 18-month period (3;3–4;9) includes a graph showing percentage inversion rates over this period – the term ‘inversion’ reflecting the theoretical position that interrogatives were derived from declaratives via the movement of the relevant auxiliary verb from a post-subject position. The graph showed a gentle upward trajectory over this period, from a 10% to about 80% inversion rate. (Which is what we would expect for just about any measure of grammatical development in this age-range). A second graph showed the breakdown of percentage inversion rates across different wh-words. A different picture emerges from this graph, depending on the wh-word. For how, where, and what questions, an 80% inversion rate applies by about 3;9. But why-questions seem to remain stubbornly resistant to pre-subject auxiliary placement until about 4;6, when they suddenly start to rise. Because the data is private to the authors of the study, we can only speculate on the reasons for this difference. Fletcher (Reference Fletcher1985) suggests that the interpretation of forms such as what's and where's, with contracted auxiliary/copula, as inverted elements, may have inflated the inversion values for these wh-words. But until the existence of a suitable body of accessible data it was not possible to take the issue further. Fast forward to volume 27 of the Journal, in which Rowland and Pine (Reference Rowland and Pine2000) take advantage of the availability of 55 hours of data in CHILDES from Adam, one of Brown's subjects, to look at the issue of inversion in his development of wh-questions. The contributions of this study, I would argue, are three-fold: descriptive, theoretical, and practical.

First, access to a database of this size, and the potential that CLAN affords for automatic analysis, forces a realistic perspective on the sheer complexity of the learning task facing the child, simply by describing the number of combinations of wh-word + auxiliary occurring in the data. The possible combinations of just four of the wh-words with twenty non-negative uncontracted copula/auxiliary forms (do does did am is are was were have has had plus nine modals) total eighty (and the figure increases if we include contracted forms). If we include negatives the number doubles. Rowland and Pine (Reference Rowland and Pine2000, p. 172) tabulate forty-nine different combinations of wh-word + auxiliary in Adam's questions. This was at a point in his development when he was using auxiliaries in these structures in both pre-subject and post-subject position – producing examples like how do you open that? as well as what he can ride in? A wide range of auxiliary forms of be, have, and do are found in the forty-nine combinations – all of those listed above, aside from were – along with six modals (can could may shall should will). Negative forms also appear: doesn't, don't, didn't, can't, couldn't. But in only two of the forty-nine wh-word + auxiliary combinations – how + can and why + is – do we find the auxiliary appearing in both pre-subject and post-subject positions. In all other cases combinations occur in one position but not in another, e.g.:

• Auxiliary pre-subject    Auxiliary post-subject

• What are    14     0

• Why + can't    0     10

Piecemeal development in relation to auxiliaries is also apparent in the elicitation studies of Ambridge, Rowland, Theakston, and Tomasello (Reference Ambridge, Rowland, Theakston and Tomasello2006), Theakston and Rowland (Reference Theakston and Rowland2009), and Rowland and Theakston (Reference Rowland and Theakston2009). The descriptive detail afforded by the analysis of large bodies of data through CHILDES, and the validation of hypotheses that arise from these studies via experimental approaches, lead to a different perspective on how the child is developing knowledge of syntax. The asymmetric character of wh + auxiliary forms suggests that an account which recognizes low-scope lexically specific knowledge of wh-word + auxiliary sequences may provide a truer perspective on this phase of syntactic development than the invocation of a movement rule. It may be that an accumulation of lexically specific sequences leads eventually to full recognition of a relationship between auxiliaries in pre- and post-subject position, but this could well be a late abstraction from accumulated examples or even a post-literacy realization.

The constructivist account, I would argue, can also contribute to the practical problem of how we think about intervention for children with language impairment. Fey and Loeb (Reference Fey and Loeb2002; see also Fey & Finestack, Reference Fey, Finestack and Schwartz2009) report on a study which was intended to improve the control of auxiliary verbs by children with specific language impairment (SLI). The question was whether children with SLI who were modelled auxiliary verbs (is and will being primary targets) in initial (inverted) position in yes/no questions would be likely to develop these and other auxiliaries. Sixteen children with SLI were involved in the research, and were subject to enrichment in the form of interrogative recasts of declarative sentences:

The children with SLI, who were on average aged 3;2, were not using auxiliaries at the time of recruitment into the study. The experimental group attended twenty sessions of enrichment over eight weeks, and during each session heard thirty recasts. There was a control group who were involved in a play condition with no systematic recasting in the same number of sessions over the same time period. There was no evidence at the end of the study that the children in the enrichment group had improved their control of auxiliaries is or will, or indeed any other auxiliary form. Fey and Finestack (Reference Fey, Finestack and Schwartz2009, p. 525) put this down to the recasts (the adult response to the child in the form of a question, as in the example above) being too complex for the child to make use of. This may be the case, and there may also be other reasons for the lack of success of the intervention. But it is also possible that the underlying assumption of the study, that there is an intrinsic relationship between auxiliaries in pre-subject and post-subject position, is not valid, at least for children with language impairment at a certain stage of development. Perhaps an intervention strategy should be guided by a different take on typical development, and start from a constructivist position – one that assumes the child gradually accumulates lexically specific sequences rather than one that takes the relationship between pre-subject and post-subject auxiliaries for granted.

A constructivist view of wh-question development is not uncontested. Westergaard (Reference Westergaard2009), using data from Norwegian and the same data from English that Rowland and Pine (Reference Rowland and Pine2000) analyze, argues that the acquisition of wh-questions must be the result of a rule-based process. While I would be more inclined to a constructivist view, my reason for seeing CHILDES as one of the most important developments in the field is not because the available data automatically favours one position over another, but because it provides an open data source which can act, among other things, as a level and public playing-field for the evaluation of competing explanatory accounts.

2. THE MACARTHUR-BATES COMMUNICATIVE DEVELOPMENT INVENTORY – MBCDI

The other milestone in child language studies which I would highlight is also a data source – the MBCDI (Fenson, Marchman, Thal, Dale, Reznick & Bates, Reference Fenson, Marchman, Thal, Dale, Reznick and Bates2007), together with its many adaptations into languages other than English (http://www.sci.sdsu.edu/cdi/). Relying on parental response to a checklist of vocabulary items and grammatical features was in the 1990s an entirely novel way of tracing language development between ages 0;8 and 2;6. By comparison with the work involved in generating and transcribing language samples, parent reports are an efficient way of collecting extensive data on large numbers of children relatively quickly, either via paper forms or, more recently, on-line (Kristoffersen et al., Reference Kristoffersen, Simonsen, Bleses, Wehberg, Jørgensen, Eiesland and Henriksen2013). The CDIs have proved to be reliable and valid instruments for monitoring early child language, especially vocabulary. Inter-child variation in rate of language development had been apparent even in the data from Adam, Eve, and Sarah, the three children in Brown (Reference Brown1973, p. 57), and was confirmed in subsequent studies based on larger samples, e.g., Wells (Reference Wells1985), which involved 128 preschool children. But the increase in numbers of children in this project had to be set against a much more limited amount of data from each child. This is a trade-off that the CDI is not subject to. With parent report, much larger samples can be accumulated, samples that are appropriately representative of a population. This enables us to see the true scope of individual variation, and facilitates the exploration of factors relevant to inter-child differences.

Stokes and Klee (Reference Stokes and Klee2009), looked at vocabulary recorded via a UK CDI (Klee & Harrison, Reference Klee and Harrison2001) from over 200 British children between 2;0 and 2;6 (mean age 26·84 months). The average number of words in the sample was 329, but the range of vocabulary items went from 15 to 666. This degree of variation in large samples fits with information from the CDI on children in the USA, and also with adaptations of this procedure in more than forty languages around the world. Similar vocabulary ranges for two-year-olds can be found for languages as typologically distinct as Irish, French, and Cantonese (O'Toole & Fletcher, Reference O'Toole and Fletcher2010; Stokes, Kern & Dos Santos, Reference Stokes, Kern and Dos Santos2012: Tardif, Fletcher, Zhang & Liang, Reference Tardif, Fletcher, Zhang and Liang2008). Stokes and Klee (Reference Stokes and Klee2009) explored the relevance of demographic, background, and processing variables to vocabulary scores, but were only able to identify one predictor which accounted for a sizeable proportion (36%) of the variance in vocabulary attainment. One of the processing tasks that the children underwent was a non-word repetition (NWR) procedure. It was the children's performance on the NWR (containing novel words from one to three syllables in length) which was the strongest predictor of vocabulary level. The child's ability to repeat novel words has been interpreted as indicative of the efficiency of the phonological loop, a component of short-term memory that allows the child to hold on to auditory material briefly before transferring it to longer-term memory. So this finding links vocabulary size in two-year-olds to a component of the child's learning mechanism. A deficiency in the phonological loop has been identified in children with specific language impairment (e.g., Gathercole, Reference Gathercole2006).

Stokes (Reference Stokes2010) has further exploited the data from this large cohort of two-year-olds by examining the effects of the neighbourhood density (ND) and word frequency (WF) of the items in the parent report on expressive vocabulary size. She found that these factors together accounted for a remarkable 61% of the variance (ND 47% and WF 14%) for the sample overall. But children with low vocabulary scores (below −1 SD of the distribution) scored significantly higher on neighbourhood density and lower on word frequency than children with high vocabulary scores. A very similar picture is seen in a comparable study on French two-year-olds (Stokes et al., Reference Stokes, Kern and Dos Santos2012). ND and WF together predict 62% of the variance in vocabulary size in a sample of over 200 French-speaking children. And the children at the low end of the French vocabulary continuum also prefer words with higher ND and lower WF values. It seems that the processes the low vocabulary children in either language are employing, in extracting and storing words from the ambient language, are restricting their development of expressive vocabulary.

These studies demonstrate the potential of CDI data. It can be mined in a way that sheds light on the learning process. In the studies by Stokes and colleagues it is similarities in learning style cross-linguistically that beg for an explanation. We can also cite examples of differences that emerge from CDIs across languages, which need to be accounted for and can lead to promising proposals. Bleses et al. (Reference Bleses, Vach, Slott, Wehberg, Thomsen, Madsen and Basbøll2008) compare the Danish CDI results to those from a number of other languages. A comparison of vocabulary comprehension shows that, for Danish children,from age 1;0 on their score is the lowest of all the languages examined. This is attributed to the characteristics of Danish phonology, which involves vowel reduction and obstruent vocalization processes, leading to potential word segmentation problems for the learner. It is striking that children learning a closely related language, Swedish, where syllable boundaries are more clearly identifiable, show a faster rate of comprehension vocabulary development from age 1;0 on than their Danish counterparts.

Tardif, Fletcher, Liang, and Kaciroti (Reference Tardif, Fletcher, Liang and Kaciroti2009) also exploit differences in CDIs between closely related languages to illuminate factors relevant to acquisition, this time in relation to grammar. Separate CDIs were normed in Hong Kong for Cantonese, and in Beijing for Mandarin. Although mutually unintelligible, these languages are closely related. Nevertheless, significant differences were found between the two populations in the speed at which they acquired vocabulary. For children in Beijing at two years of age, the median vocabulary size was 657, while for children in Hong Kong at the same age, it was 322. Some of this discrepancy seems to be due to the single child status of children in Beijing, and also to their monolingual language environment. This is, however, not the full story, which may well, as with the Danish/Swedish comparison, involve phonological factors (Tardif et al., Reference Tardif, Fletcher, Liang and Kaciroti2009, pp. 1139–1140). The interest here is in a specific grammatical category common to the two languages, that of classifiers. If we simply look at number of classifiers used by age, we find the overall difference in vocabulary score reflected in this value: two-year-old Mandarin speakers – and indeed children at each month up to 2;6 – are reported to use significantly more classifiers than their Cantonese-speaking counterparts. But if we equate children by vocabulary size, this difference reverses, and the Cantonese-speaking children have the advantage. This can be shown for the grammatical category of classifiers overall, and for individual classifier types. So from a vocabulary size of 200 words on, Cantonese-speaking children have the advantage over their Beijing counterparts. What we see revealed here, it can be argued, is the attentiveness young children pay to subtle features of the input. In Cantonese, classifiers are obligatory in more grammatical contexts than in Mandarin, and a wider range of semantically specific classifiers (as opposed to the general classifier, which exists in both languages) tends to be used in the speech of Cantonese adults. We may surmise that it is these input factors that encourage Cantonese-speaking children to tune into the precise functions and extensions of classifiers more readily than their Mandarin-speaking counterparts.

CHILDES and the CDI then both contribute in significant ways to the availability of data in the field, within and across languages. CHILDES is an open data source, primarily for language samples. CLEX (Jørgensen, Dale, Bleses & Fenson, Reference Jørgensen, Dale, Bleses and Fenson2010) continues to extend the range of languages on which CDI data is available, with ten now represented. CLEX is a website available to all, that allows researchers to query all aspects of the CDI dataset or sets they are interested in, construct tables and graphs from the data, and make comparisons across languages. The benefits CHILDES and the CDIs bring to the field of child language in terms of data provision are substantial, and represent a dramatic change to the situation at the time of this Journal's foundation. But as I hope I have sketched out above, the contribution they make to our profiling of the course of language development is considerable, and their impact on the understanding of the process of language learning is not negligible.