Negative Input for Grammatical Errors:

Effects of negative input for 13 categories of grammatical error were assessed in a longitudinal study of naturalistic adult-child discourse. Two-hour samples of conversational interaction were obtained at two points in time, separated by a lag of 12 weeks, for 12 children (mean age 2;0 at the start). The data were interpreted within the framework offered by Saxton’s (1997; 2000) contrast theory of negative input. Corrective input was associated with subsequent improvements in the grammaticality of child speech for three of the target structures. No effects were found for two forms of positive input: non-contingent models, where the adult produces target structures in non-error-contingent contexts; and contingent models, where grammatical forms follow grammatical child usages. The findings lend support to the view that, in some cases at least, the structure of adult-child discourse yields information on the bounds of grammaticality for the language-learning child. Negative Input for Grammatical Errors: Effects After a Lag of 12 Weeks Definitions of Negative Input Do parents correct their children’s grammatical errors? For many researchers, Brown & Hanlon (1970) settled this question more than three decades ago when they reported that parental signals of Approval and Disapproval are not contingent on the syntactic wellformedness of child speech. It appeared that grammatical deviations were allowed to pass unchecked, leaving children to their own devices in determining the crucial distinction between what is grammatical and what is not grammatical. This finding generated huge interest, even being heralded as ‘one of the most important discoveries in the history of psychology’ (Pinker, 1988:104). While this may be overstating the case, Brown & Hanlon’s findings did seem to provide empirical support for Chomsky’s (1980) argument from the poverty of the stimulus. Normally-developing children eventually retreat from error to attain a mature system of grammar. If one cannot explain the origins of their grammatical knowledge in terms of support from the linguistic environment, then one can conclude instead that this knowledge is innate. Parental correction comprises a form of negative input, that is, information that an utterance is ungrammatical. If available, negative input could function as a powerful constraint on language acquisition, since it conveys the precise scope of grammatical rules to the child. In many cases this entails the retreat from overgeneralization, where the child grammar is a superset of the adult grammar. A well-attested example is the phase where children make errors with past tense forms of irregular verbs (e.g., Marcus, Pinker, Ullman, Hollander, Rosen, & Xu, 1992). During this phase, the child grammar permits both breaked and broke, while the adult grammar is more restricted in allowing only broke. Negative input could also be useful in underscoring the obligatory nature of certain forms. Thus, children pass through a phase where structures that are obligatory in the adult grammar appear to be optional in the child's grammar. A case in point is the omission of obligatory morphemes like determiners, plural suffixes and regular past tense markers. It is possible to view optional omissions as a form of overgeneralization, since the child grammar permits two forms (e.g., dog and dogs) where the adult grammar allows only one (dogs). The plausibility of this view is lent credence by the fact that many children with specific language impairment persist for years, even indefinitely, with optional omissions of this kind (e.g., Rice, Wexler & Cleave, 1995). Whether they are viewed as a form of overgeneralization or not, it is clear that negative input could be useful in confirming the obligatory nature of particular forms. It cannot, however, specify the nature of what is acquired in the first instance, nor how it is learned. Negative input is confined, instead, to the process of 'unlearning', or rather, the shedding of ungrammatical forms from the child's nascent grammar (Saxton, 1997). The Contrast Theory of Negative Input The power of the argument from the poverty of the stimulus depends on the extent to which children are supplied with negative input. In this respect, much empirical evidence has accrued since the 1980s to indicate that Brown & Hanlon (1970) were, in all likelihood, premature in their conclusions (e.g., Hirsh-Pasek, Treiman & Schneiderman, 1984; Demetras, Post & Snow, 1986; Bohannon & Stanowicz, 1988; Farrar, 1992; Saxton, 2000; Strapp & Federico, 2000; Chouinard & Clark, 2003). All of these studies report that the input to young children is replete with responses to grammatical errors which, prima facie, look like corrections. A fundamental challenge is to demonstrate that what might look like a correction to the adult eye is, in fact, interpreted in that way by children from the earliest stages of grammatical development. One approach to this problem is to examine children's responses to putative sources of correction. One can be more confident that an adult response is corrective, if children utilize such input in the retreat from error. However, much of the empirical work on corrective input has failed to address these issues, leading, in certain cases, to conceptual difficulties in the interpretation of findings (see Saxton, 1997; in press, for reviews). A recent exception to this trend, however, has been the Direct Contrast hypothesis advanced by Saxton (1997, 2000) to explain how key adult responses (negative evidence) might fulfil a corrective potential. On this approach, negative evidence is identified in cases when an erroneous child form is followed directly by the correct adult alternative as in the following three examples. 1. Child: I thought they were all womans. Adult: They’re not all women. 2. Child: It’s bored of being on the bike. Adult: It’s not boring. 3. Child: He’s got little nice feet. Adult: Oh, he has got nice little feet. The Direct Contrast hypothesis predicts that the corrective power of negative evidence lies in the immediate juxtaposition of child error and correct adult alternative (Saxton, 1997). It is predicted that the contrast between the two forms is rendered especially salient in this particular discourse context. In (1), both child and adult are focused jointly on the same topic of conversation, and both make use of a plural form of woman. Yet the adult conspicuously eschews the child’s selection of womans in favour of women, and the contrast in usage between the two forms is thrown into sharp relief. Negative evidence creates a discourse context that is unique in its power to reveal the contrast in usage between alternative forms. In making this contrast, it is predicted that the child is supplied with two prerequisites, without which the retreat from error would not be possible. First, the child needs to know that the form favoured by the adult is grammatical. And second, the child needs to be apprised of the fact that their own selection is ungrammatical. Observe that information about the former point is available in abundance. Adults cannot speak to a child without modelling grammatical forms, in both error-contingent and non-error-contingent contexts. These linguistic forms, modelled by the adult, are traditionally referred to as positive input (e.g., Crain & Pietroski, 2001; but see Clark, 2003, for an alternative approach couched in terms of her Principle of Conventionality). There is a general consensus that positive input alone cannot explain the child’s retreat from error. One reason is that simply being apprised of the grammaticality of women, as in (1), does not, in and of itself, constitute evidence about the ungrammaticality of womans. Of importance here is the fact that the young child’s grammar is characterized by overgeneralization. That is, it permits both grammatical and ungrammatical forms at the same time, often for protracted periods extending into months and even years (e.g., Marcus et al., 1992). Hence, it is not sufficient simply to be apprised of the fact that women is grammatical. The child already knows this. The critical piece in the puzzle is information concerning the ungrammaticality of womans. The Direct Contrast hypothesis assumes that this information is embedded in the structure of the discourse. The juxtaposition of child and adult forms is held to supply information over and above the simple existence of women as an acceptable adult form. In creating a contrast in usage between the two alternatives within the discourse, it is predicted that the ungrammaticality of womans is revealed to the child. In according the errorcontingent adult model this special status, the contrast theory diverges sharply from traditional linguistic approaches to child language acquisition (e.g., Crain & Pietroski, 2001). In the latter case, the discourse structure of adult-child conversation is irrelevant. Grammatical forms modelled by the adult are taken to have the same status, regardless of their contingency on grammatical errors. Recent evidence on the status of error-contingent models is reviewed below. A critical consideration is how the child might alight exclusively on points of contrast relevant for the acquisition of grammar. Saxton (1997) suggests that the equivalence in grammatical function of, say, women and womans provides the key to this issue. Typically, the child's overgeneralized system already exemplifies, in this case, both women and womans, so the equivalence in function between the two is known in advance to the child. Given this prior knowledge, the child needs to know that one of the forms they allow (womans) is in fact ungrammatical. The Direct Contrast hypothesis predicts that errorcontingent adult models hold a special status in this regard, since they have the power to alert the child to the contrast in usage between child and adult speech. Beyond the issue of functional equivalence, there is often a significant overlap in phonological form between erroneous and correct forms. Such phonological overlap could well prove beneficial in alerting the child to relevant points of contrast, although the equivalence of grammatical function remains the driving force in triggering retreat from error. An important assumption in the preceding discussion is that the child interprets only relevant points of contrast within adult-child discourse as a form of negative input. Adult responses to child utterances contrast in an infinite number of ways. What, then, prevents the child from focusing on spurious points of contrast and rejecting perfectly acceptable forms in consequence? 4. Child: Pretend I dropped my bow and arrow down. Adult: You dropped it on the floor? If the child were identifying points of contrast at random, they might substitute the adult's floor for their own form down in (4) above. But this is unlikely. Acquisition would be nigh on impossible if the child misconstrued every adult utterance in this way. Of course, this kind of destabilisation is not empirically attested. In the first instance, it is difficult to conceive what might motivate such random substitutions. Certainly, it could not be the equivalence of grammatical functions alluded to above. Unless there is good reason to suppose that the child might perceive this kind of equivalence, there will be nothing to trigger the perception of a direct contrast by the child. Other forms of contrast between child and adult utterances may be more problematic. One such contrast occurs in (5) below: 5. Child: The alligators swim. Adult: The alligators will swim. In this (constructed) exchange, an ostensibly grammatical child utterance is met with a highly similar, but nevertheless contrasting, adult response. However, Saxton (1997) points out that the child utterance here has at least two possible interpretations. The alligators swim may refer to an habitual event, in which case it is highly unlikely that the adult would respond as here. Or the child may be referring to a future event, in which case The alligators swim can be construed as ungrammatical on the grounds that an obligatory auxiliary verb (will) is missing. In this case, the adult response can be construed as negative evidence, on the definition and interpretation offered within the contrast theory (see Saxton, 1997: 157-159 for more detailed discussion). A further potentially damaging kind of contrast is offered by a reviewer in the constructed exchange reproduced in (6) below: 6. Child: I have a dog. Mother: I have two dogs. At first blush, this example seems to stretch credulity somewhat. In the first instance, we have to strain to contemplate a family where the mother owns two dogs quite separately from her two-year-old child, who owns just one, and where it is news to both parties that either owns any kind of pet at all. But maybe mother and child are talking about toy dogs, or pictures of dogs. If so, then we are left with an ostensibly grammatical child utterance, which, in tandem with the adult response, yields a contrast between the two forms dog and dogs. If the child perceived a direct contrast in this case, they might reject the use of dog in this context, with future utterances being more like I have a dogs. On this scenario, then, the child’s understanding of plural markers would be in serious peril. In fact, though, the Direct Contrast hypothesis does not apply, because there is no basis for the child to apprehend any functional equivalence, grammatically, between the child and adult forms. In the first instance, the child seems to be talking about one object, while the adult is talking about two. Not only do both speakers mark the nouns appropriately, they also, incidentally, use appropriate modifiers. Moreover, it is likely in this particular case that the adult would place contrastive stress on two, thus highlighting in a very explicit manner the intention to talk about something different from the child: two objects instead of one. The distinctiveness of the grammatical functions being expressed would, in all probability, also be underscored by the context of utterance. In short, it is difficult to see what might mislead the child into apprehending the adult use of dogs as an alternative for talking about a single dog. Of course, children often omit the plural marker –s when referring to more than one object. If the child in (6) intended to express plurality, then, they might perceive the adult use of dogs as contrasting with their own. One would need to demonstrate that both child and adult were referring to two (or more) dogs, at the same time, in the same context of utterance, in order for the child to apprehend that the adult usage contrasted with their own. The key, then, in all cases is to determine whether the child could plausibly apprehend two linguistic forms as fulfilling an identical grammatical function, in a given context of utterance. Without this prior cognisance, there is nothing, on this view, that could trigger the rejection of their own form for the adult alternative. It may seem trivial to consider constructed examples of this kind in such detail. If the examples in (5) and (6) do not serve the discussion well, then perhaps one might readily find more suitable alternatives. We would argue, however, that it is not at all easy to come up with credible fictions in the domain of adult-child discourse. The very nature of the impulse to find extreme examples that disprove a theory often renders them inherently more likely to lack empirical plausibility. Even where such cases are logically possible, unless they are attested in genuine interactions, their value in testing the theory is diminished. When one examines actual child-adult discourse, there is no evidence yet forthcoming that the child is misled by spurious points of contrast within the discourse. Moreover, the Direct Contrast hypothesis provides a plausible explanation for why the child will only ever focus on relevant contrasts. Availability of Negative Input Something in the region of twenty studies have now demonstrated that negative evidence, as defined here, is supplied to young children (see, for example, Demetras et al., 1986; Bohannon & Stanowicz, 1988; Farrar, 1992; Saxton, 2000; Strapp & Federico, 2000; Chouinard & Clark, 2003). Of immediate note, though, is the wide variation in levels of corrective information reported. With respect to negative evidence, Farrar (1992) reports a frequency of 3% for past tense errors compared with 44% for article errors. More recently, Chouinard & Clark (2003) report levels of negative evidence as high as 65% in some instances. A number of factors underpin the wide range in levels of negative evidence reported. These include individual differences among parents and the child’s other interlocutors, the linguistic level of the child, and the particular grammatical structure under review. The linguistic level of the child is important if only because, as development proceeds, the child produces fewer and fewer grammatical errors. Hence, opportunities to provide corrections progressively diminish. With regard to the particular structure, it is not clear why certain categories, like the past tense errors in Farrar’s study, should be subject to relatively low levels of correction. In other cases, though, the level of corrective input witnessed is more easily explained. For example, Saxton (2000) reports that 51% of subject errors meet with negative evidence. But given that adults can scarcely produce a full sentence without a grammatical subject in English, it is hardly surprising that a high level of child subject omissions are followed by adult responses which model a sentential subject. Despite considerable variation in the frequencies reported, negative evidence has been reported in every empirical study on the topic. In fact, for every individual child for whom data are available, and also for every grammatical structure examined, negative input, as defined here, is supplied. The issue of availability can be extended to consider whether corrective input is available for every single child. The answer to this question is of theoretical interest, if only because negative input could be dismissed as a necessary component of language development if even one child could be identified that had been denied access to corrective information. Unfortunately, cross-cultural research on this issue is extremely scarce. What little there is tends to lack empirical rigor, relying instead on anecdotal descriptions. An exception is provided by Chouinard & Clark (2003) who report the occurrence of negative evidence for French-speaking children. A more sceptical view is hinted at by Ochs (1982) in her discussion of Western Samoan society. She reports that parents spend little time interacting directly with very young children and provide few of the features of child directed speech familiar in Western settings. However, in the absence of more detailed empirical evidence, two notes of caution should be sounded. First, in Samoan society, the task of interacting with very young children is commonly assigned to older siblings. Given that children as young as 4;0 produce the modifications associated with child directed speech (Shatz & Gelman, 1973), the possibility is open that older siblings may provide corrective input in Western Samoa. Beyond that, Ochs (1982:101) observes that parents do in fact paraphrase their children's speech. No category definition is supplied, but it is plausible that paraphrasing might encompass the kinds of error-contingent adult models investigated here (see Chouinard & Clark, 2003, for further discussion). These remarks underscore the fact that cross-cultural research is needed that directly addresses the issue of negative input. At the very least, it should be apparent that conclusions about the general non-occurrence of negative input are, as yet, premature. Effects of Negative Input Having established that both error-contingent models and error-contingent clarification questions are available to the child, it behoves researchers to assess what impact, if any, they have on grammatical development. In this regard, researchers have begun to investigate both the immediate and longer-term effects of (potentially) corrective input. With respect to immediate effects, a number of studies report a propensity on the part of children to reject erroneous forms and switch to the grammatical counterpart modelled for them by the adult (Farrar, 1992; Morgan, Bonamo & Travis, 1995; Saxton, 1997; 2000; Strapp & Federico, 2000; Chouinard & Clark, 2003). As with frequency of negative input, the rates at which children pick up on corrective information in this way varies considerably. Shifts in child speech from erroneous to correct (E → C) vary between 8% (Saxton, 2000) and 30% (Saxton, 1997), with levels for individual children reaching 58% (Morgan et al., 1995) and 45% for individual grammatical structures (Farrar, 1992). The frequency of corrective input may well be significant, in that it may be necessary to exceed a particular threshold level in order to exert any influence on the child. At present, there is very little evidence about what such threshold levels might be, nor yet whether children vary in their receptiveness to negative input. An indication is provided by the experimental studies conducted by Saxton and colleagues (1997; 1998). In these studies, some children rejected erroneous forms in favour of the correct version modelled for them after only one exposure. It is not clear, however, how much corrective input is needed to effect permanent improvements in the child's grammar. Turning to longer-term effects, data from experimental and intervention settings all suggest that the effects of negative evidence can be observed after lags ranging from several weeks to several months (e.g., Saxton, Kulcsar, Marshall & Rupra, 1998; Proctor-Williams, Fey & Loeb, 2001). No effect was found, however, when Morgan et al. (1995) applied the econometric method of time series analysis to the data in Brown’s (1973) corpus on three children. However, objections have been raised concerning both the adequacy of the modelling procedures used (Bohannon, Padgett, Nelson & Mark, 1996) and to the suitability of Brown’s data for time series analysis (Saxton et al. 1998). On the latter point, the data in Morgan et al.’s study seriously violated the requirement that data be gathered at strictly regular intervals (for detailed discussion, see Saxton et al., 1998). In examining the effects of negative input, it is not sufficient merely to examine improvements in the grammaticality of child speech. If the nativist description of the input is correct, what is described as negative evidence here would count as a subset of positive input. One could therefore attribute any effects witnessed to positive, rather than negative, input. One way of tackling this problem is to compare the effects of error-contingent and non-errorcontingent models. On the nativist view, the context in which forms are modelled should make no difference in the effects observed on the child. In the contrast theory, on the other hand, the difference between the two is critical. In fact, available evidence supports the contrast theory. First, error-contingent models encourage significantly higher levels of E→C switches than non-error-contingent models (Farrar, 1992; Saxton, 1997; Saxton et al., 1998; Saxton, 2000;). Second, E → C shifts are more frequent following negative input than following adult ‘move-ons’ (no correction; Morgan et al., 1995; Saxton, 2000). And third, children’s intuitions concerning the status of grammatical and ungrammatical forms are closer to adult intuitions when past tense forms are modelled as negative, rather than positive, input. Overall, therefore, evidence is emerging that error-contingent models enjoy a special status in the input to the child, supplying a richer source of information on grammaticality than their non-error-contingent counterparts. In the current study, we examine two types of positive input to provide bases for comparison with the effects of negative evidence. The first category, henceforth referred to as contingent models, comprises all adult models of target structures contingent on grammatical child utterances. The second, referred to here as non-contingent models, comprises all nonerror-contingent adult models of target structures. Evidently, the first category comprises a subset of the second, and is promulgated on the assumption that contingent adult models may reinforce correct child uses and hence be of especial benefit to the child. The current study aims to broaden and extend recent findings within the framework offered by the contrast theory. Data are currently required on the longer-term effects of negative evidence for normally developing children in naturalistic conversational settings. Studies with SLI children (e.g., Proctor-Williams et al., 2001) and experimental data (Saxton et al., 1998) lack ecological validity when it comes to assessing the effects of corrective input in the course of normal language acquisition. Without this kind of data, the relevance of negative input for theories of child language acquisition cannot be adequately assessed. Hence, the current study examines the effects of negative input on child speech after a lag of 12 weeks. Morgan et al. (1995) notwithstanding, the only extant data of this kind are reported by Proctor-Williams et al. (2001) for their control group of typically developing children with respect to two aspects of grammar, articles and copulas. Negative evidence was associated with improvements in the speech of children aged 2;0 for one of these structures (copulas) after a lag of eight months. The present study seeks to extend both the range of structures investigated and the size of language samples examined. In addition, the effects of negative evidence will be compared against the effects of the two forms of positive input described above (contingent and non-contingent models). Method Design A correlational design was implemented, with frequencies of negative input at Time 1 being correlated with child grammaticality (CG) at Time 2. At first glance, standard concerns about direction of causation seem unnecessary, by dint of the lag between Time 1 (T1) and Time 2 (T2). It seems reasonable to assume that negative input at T1 might influence child grammaticality at T2, but not vice versa. However, it has long been established that the influence of adult input on child speech is tempered by the countervailing influence of the child’s language on the speech of adults. In fact, child grammaticality at Time 1 (CG1) may be the causal factor underlying any positive correlation between negative input at T1 (NI1) and child grammaticality at T2 (CG2). This situation arises where child grammaticality at T1 underpins child grammaticality at T2, while at the same time being the causal influence on the adult’s corrective behaviour at T1. Given these caveats, an initial precaution is that children should be as similar as possible with respect to their language level at T1 (e.g., Scarborough & Wyckoff, 1986). A further consideration is the need to measure gains in child grammaticality from Time 1 to Time 2 (Richards, 1994). One reason is that a positive correlation might arise between NI1 and CG2, even in the absence of such gains. In this instance, any causal relationship between negative input and child grammaticality would be short-term via the child at T1. Unfortunately, gain scores tend to be negatively correlated with initial child status. Children with high scores for grammaticality at T1 are likely to make relatively small gains, while low scorers at T1 are likely to make larger gains. This diminution in variance has the consequence of attenuating the NI1xCG2 correlation. There is a danger, then, of producing conservative results which could underestimate the true strength of the NI1xCG2 relationship. The preceding discussion serves to establish that simple gain scores for child grammaticality are best avoided. At the same time, Richard’s (1994) observation that gain scores are required to gauge long-term effects remains valid. One solution to this problem, advocated by Richards and adopted here, is to use the residual gain scores derived from multiple regression analyses, since they have a zero correlation with CG1 by definition. In consequence, one removes what Gleitman, Newport & Gleitman (1984) refer to as ‘the effects by the child on the child’ (p.46). One is left, however, with ‘the effects of the child on the mother’ (ibid., p.46), in this case, the effect of CG1 on NI1. This latter problem can be addressed by partialing out the effects of CG1 from the NI1xCG2 correlation. This procedure allows one to use raw, rather than gain scores, since it yields the equivalent of a residual gain score (removing the effect of CG1 on gains in grammaticality at T2). At the same time, of course, the effect of the child on levels of negative input at T1 are removed. In summary, both regression analyses and partial correlations are used here to examine the effects of negative input on child grammaticality after a lag of 12 weeks. The lag of 12 weeks between T1 and T2 was chosen for two main reasons. Firstly, this length of time is sufficient for the child to have made appreciable gains in grammatical development (cf., Saxton, 2000). And secondly, by taking the second sample before the child reaches full mastery, ceiling effects can be avoided. More generally, many studies which have investigated other aspects of input also report clear effects after an interval of three months or thereabouts (e.g., Nelson, Denninger, Bonvillian, Kaplan & Baker, 1984). Participants Twelve mother-child pairs were recruited from nurseries in west London, UK, and came from middle class families where English was the only language spoken in the home. There were 7 boys and 5 girls with a mean age of 2;0 at the start of the study (range 1;9 to 2;5). The mean length of utterance (MLU), averaged across all 12 children, was 2.31 at the start (range 1.38 3.06). Materials Both audio and video recordings were made of the mother-child interactions. Video recordings were made on a Panasonic Super VHS AG-450 which has a relatively high quality built-in microphone. Video data were supplemented with audio recordings, made on a Sony Digital Audio Tape Corder (TCD-D8). Procedure Participants were visited in their own homes on two consecutive days at both T1 and T2. On each visit, a total of two hours of recordings were made. Owing to child boredom and fatigue, and the intervention of daily routines, each two hour sample was compiled from a number of shorter recording sessions across the two-day period. Recordings were typically made in the living room of participants’ homes, with the camera mounted on a low tripod in the corner of the room and operated by one of the authors. Participants were encouraged to engage in a range of normal activities, which included playing with toys, drawing, cooking and reading together. The lag of 12 weeks between T1 and T2 recordings was adhered to as closely as possible. In the event, recordings were made within a range of 12 weeks +/9 days. Transcription and Coding Initially, the data were transcribed and coded by the second author. Two further examiners then checked the resultant transcripts for accuracy, using the audio and videotapes. Disagreements were noted and where possible resolved by mutual agreement among examiners. Outstanding disagreements were marked as untranscribable and omitted from analyses. Utterances containing any portion of untranscribable material were likewise not analyzed further. For the purposes of analysis, each sample was confined to the first two hours of transcribed material. Child utterances were coded for both grammatical and ungrammatical uses of 13 structures: subject; object; 3rd person singular; determiner; present progressive auxiliary verb; prepositions; possessive; copula; auxiliary verbs (comprising all auxiliaries other than present progressive auxiliary); regular plural; irregular plural; regular past tense; and irregular past tense. Examples of child grammatical errors from the transcripts are provided in Table 1. It should be borne in mind that errors of omission predominate in this phase of development. It is also worth noting that child utterances often exemplify more than one kind of error, a fact which is apparent from some of the examples in Table 1. The grammatical categories investigated in this study closely mirror those adopted in Farrar (1992) and Saxton (2000). An innovation here is the creation of regular and irregular categories from the more generic past tense and plural categories. In most cases, identifying an error for a particular category was relatively straightforward, although it was necessary on occasion to refer to the wider context of utterance in order to confirm coding decisions. TABLE ONE ABOUT HERE Maternal input was coded for the following three categories: negative evidence (NE); contingent models; and non-contingent models. Negative evidence was identified according to the category definition given in (7) below. Examples from the transcripts, with respect to specific categories of error, are also given in (7). As in Table 1, errors of omission are marked with ‘^’. 7. Negative Evidence Negative evidence occurs directly contingent on a child grammatical error, and is characterised by an immediate contrast between the child error and a correct alternative to the error, as supplied by the child's interlocutor. a. Determiner Child: We haven’t got ^ box. Adult: No, we haven’t got a box. b. Irregular past Child: I drawed it for you. Adult: You drew it for me Adult utterances were also coded for positive input with respect to the thirteen target structures. Two forms of positive input were examined: contingent and non-contingent models. Non-contingent models comprised all those instances where a target structure was supplied in an adult utterance which modelled that structure, excluding all cases of negative evidence, as exemplified in (8) below. Contingent models, meanwhile, comprised instances where the adult modelled target structures contingent on a grammatical child utterance, as in (9). 8. Non-Contingent Models a. Subject, Object, Auxiliary, Regular Past Adult: Have you tried it? b. Subject, Copula, Determiner, 3rd Person Singular Adult: Where’s the biscuit? 9. Contingent Models a. Article Child: A table. Adult: Yeah we'll have a little table here. b. Subject, Regular Past, Preposition, Article. Child: I played in the sand. Adult: Have you played in the sand? The transcripts were produced in a machine-readable format compatible with the conventions of the Child Language Data Exchange System (CHILDES). CHILDES provides software for the automatic analysis of transcripts known collectively as Computerized Language Analysis or CLAN. For the purposes of the current study, CLAN allowed mean length of utterance to be calculated automatically. In addition, once the transcripts had been coded, frequency counts could also be made automatically. To assess the reliability of input coding, eight hours of material taken from four children at T1 and T2 were recoded. Given that the final corpus used in analyses comprised 40 hours of material (see Results section below), coding reliability was estimated from 20% of the total data set. For each of the three input categories, reliability was estimated by calculating the number of agreements between the two coders divided by the total number of coding judgments made (cf., Proctor-Williams et al., 2001). For negative evidence there was 93.9% agreement (339/361), for non-contingent models there was 95.7% agreement (6864/7175), while for contingent models there was 96.5% agreement (2974/3083).