Constructivism Chapter – Ambridge & Lieven
The aim of this chapter is to outline a constructivist account of the process of language acquisition, and to summarize the supporting evidence for this account, drawing on examples from some of the most intensively studied acquisition domains. Our goal is not to outline a generalized historical constructivist account, but rather to begin to sketch a new account that, in some small but significant ways, departs from previous proposals. In other words, while the account that we will outline here of course owes a considerable debt to earlier constructivist accounts (e.g., Bates & Macwhinney, 1982; Pine & Lieven, 1993; Langacker, 2000; Tomasello, 2003; Dąbrowska, 2004; Goldberg, 2006) we are speaking for no one but ourselves. We do not, in general, compare this account against rival theoretical approaches (c.f., Ambridge & Lieven, 2011; Ambridge, Pine & Lieven, in press), which we mention only very briefly, purely for comparative purposes.
That said, the account that is presented here is probably best understood with the aid of just a little historical context. Since at least Chomsky (1957), the dominant view of language acquisition has been one under which children have innate knowledge of linguistic categories and phrases (e.g., [VERB], [NOUN], [VERB PHRASE], [NOUN PHRASE]) and some language-general rules for combining them into phrases (e.g., a [VERB PHRASE] contains either a [VERB] followed by a [NOUN PHRASE] or vice-versa - e.g., [kick] [the ball] / [the ball] [kick] – with each language committing itself exclusively to one of the two possible orders).
The constructivist approach, which dates back to at least Braine (1963), arose primarily as a challenge to such claims. The basic idea is that children’s very earliest linguistic representations are not adultlike categories and rules (e.g., [VP] = [V, NP]), but rote-learned concrete holophrases (I+want+it) and low-level, lexically-specific slot-and-frame patterns or schemas (e.g., I’m [X]ing it). Only gradually do children abstract across these holophrases and lexical schemas to arrive at adultlike fully-abstract constructions. The constructivist approach is emergentist in two senses. First, it is emergentist in the sense that the generalizations that underlie linguistic competence emerge from the analysis of linguistic units stored in memory (initially, rote-learned holophrases), rather than being innately specified (as under many rival accounts). Second, the approach is emergentist in the sense that children’s language acquisition is emergent from – indeed, a byproduct of - their use of language as a social tool. Children are not “trying” to learn syntax; they are not conducting formal analyses of linguistic structure, combining content-free algebraic symbols, setting parameters, or building abstract linguistic categories for their own sake; they are using language, to cajole, to control and to communicate.
Presumably as a reaction to the prevailing claim of very early abstract knowledge, most research conducted within the constructivist framework has focused on demonstrating that young children’s knowledge is lexically specific (see Tomasello, 2000; 2003 for reviews). As a consequence, the constructivist approach has often been interpreted – by both its critics and its advocates – as claiming that, until some relatively advanced age (perhaps around thee years) all or most of children’s knowledge consists of rote-learned holophrases and lexical schemas, with any demonstration of earlier abstract knowledge taken as evidence against the approach.
In our view, this is a misinterpretation. The central claim of the constructivist approach relates not to age - “children do not have abstract knowledge until age X” - but to process: Children start out with holophrases, which develop, via a process of abstraction, first into lexical schemas, and, finally, into adultlike abstract constructions. Importantly, this process, whilst protracted and gradual, begins as soon as children have, in principle, two stored exemplars across which to abstract.
Thus, early abstract knowledge does not falsify the constructivist account: Any abstract knowledge could, in principle, have been arrived at via a process of abstraction across stored exemplars, rather than having been present all along, regardless of the age of the child. Lest this claim seem too strong, it should be borne in mind that a child who can relate teddy to a picture of a teddy in a book, to her own teddy and to a bear in the zoo has already made an abstraction; and studies with newborn infants suggest that some phonological abstractions are formed in utero (e.g., Moon, Cooper & Fifer, 1993).
Nevertheless, the constructivist account does make an eminently falsifiable prediction. Because the process outlined above is input driven, and because children’s input (and uptake) is uneven, so children’s knowledge is predicted to be uneven, in ways that correspond systematically to the language to which they are exposed. In more concrete terms, the prediction is that children will show better linguistic performance (in whatever task), when they are able to make use of a string (I+want+it) or lexical schema (e.g., I’m [X]ing it), that they have frequently encountered and thus stored in memory. Children will show worse performance on an equivalent utterance for which no stored string or template is available, even if that utterance is formally identical when analyzed at the level of adult linguistic categories (e.g., John kissed Sue, which – like I want it or I’m eating it – can be analyzed as having the structure [NPSUBJECT] [VP [V] [NPOBJECT]]). Furthermore, even when children have formed adultlike abstract constructions, they will show an advantage for utterances that constitute prototypical instances of those constructions.
On our reading of the literature, these predictions have yet to be falsified, and, indeed, enjoy considerable empirical support. In this chapter, we summarize our constructivist account of development, and the current state of the empirical evidence, for each of four particularly well-studied domains: the acquisition of (i) determiners, (ii) inflectional morphology (iii) basic word order and (iv) more advanced constructions (datives, locatives, passives, questions, and relative-/complement-clause constructions).
We by considering one of the smallest and most restricted linguistic domains: the English determiner system. Setting aside, for a moment, both the pragmatic aspects of the system and more borderline category members, all children have to learn is that English has two determiners, the and a(n), and that – on the whole - if a particular noun has appeared with one determiner it can appear with the other (e.g., the ball, a man; the book, a man etc.). The highly restricted nature of this system means that it constitutes both an excellent example with which to illustrate the constructivist account, and a popular test case for this approach.
The constructivist account of the acquisition of this system runs as follows. Suppose a child hears, and stores, the following strings:
a ball the ball
a book the book
a doggie the rain
a man the pen
The child will schematize across the strings in the first column to form the lexically-specific slot-and-frame schema a [X] and across the strings in the second column to form the schema the [Y]. This process is outlined in detail later in this section. For now, the important point is simply that these schemas allow children to produce determiner+noun combinations that they have never heard before. For example, a child who had heard a man but not the man could produce this latter combination by inserting man into her the [Y] schema.
Because the schematization process is slow and gradual, there will be a point early in development in which these slot-and-frame schemas are not yet fully formed, with children relying – at least some of the time – on the use of rote-learned strings (e.g., a+man; the+rain). Thus the constructivist account makes a simple prediction: If we can catch children at this very early stage, there will be some nouns that appear in their speech with a and not the – and vice-versa - because only the former has been stored as part of a rote-learned string (e.g., the child has stored a+man but not the+man), and a productive schema that could be used to generate it (e.g., the [Y]) has not yet been formed. Of course, in any given sample of adult speech, some nouns will be used with a and not the, or vice versa, simply for discourse reasons (e.g., the phrase a drink is used much more frequently - often with Do you want… - than the drink). So the prediction is not simply that children’s overlap between the and a uses of a particular noun will be low – the same is true for adults – but that this overlap will be significantly lower for children than adults (i.e., their caregivers).
Precisely how to test this prediction fairly has been the subject of a long-running methodological debate (Pine & Martindale, 1996; Pine & Lieven, 1997; Valian, Solt and Stewart, 2009; Yang, 2010; Pine, Freudenthal, Krajewski & Gobet, 2013). The upshot is that it is important to restrict the analysis to nouns that (a) can combine grammatically with both the and a(n) (e.g., *an advice), (b) are used at least twice by a given speaker (hence giving the potential opportunity for overlap to be observed) and (c) are used by both a given child and his caregiver (otherwise, adult overlap rates are artificially depressed by low frequency nouns that have little opportunity to appear with both the and a, and which children do not use). When this is done (Pine et al, 2013), naturalistic data studies reveal a significantly lower overlap rate for children (31%) than their caregivers (47%); a finding that, incidentally, constitutes evidence against rival accounts under which both the determiner and noun categories, as well as some knowledge of how to combine them, are present from birth (e.g., Valian, 1986).
However, as we stressed in the introduction, abstract knowledge is not all-or-nothing, and these findings do not demonstrate that young children are relying entirely on rote-learned determiner+noun strings. Indeed, Pine and Martindale (1996) argued that children showed evidence of having acquired some low-level slot-and-frame schemas (e.g., That’s a [THING]; On the [SURFACE]) which, despite their rather contextually-specific nature, do enable at least some nouns (e.g., table, chair) to be used with both a and the.
From rote-learned phrases to lexically-specific schemas
As well as an important test case for the constructivist account, the English determiner system is useful as an example of the process of schematization assumed by this account. Returning to the example above, suppose that the child has stored the strings a ball, a book, a doggie, a man, the ball, the book, the rain and the pen. The child then schematizes across the first four strings to form an a [X] schema, and across the last four strings to form a the [Y] schema.
The use of X and Y to denote the slots is particularly important for two reasons. First, we have avoided using terms that relate to adult categories (e.g., [NOUN]) in order to emphasize the claim that children have not formed such categories (indeed, we suggest below that they may never do so). Second, we have avoided using a generic term to label both slots (e.g., [THING]), in order to emphasize the claim that the [X] and [Y] slots have different, though overlapping, properties.
What does it mean for a slot to have a property? The property of a slot is a weighted average of all the items that have appeared in this position in the input utterances that gave rise to the schema. So, for this artificially restricted example, the property of the [X] slot will be a weighted average of the properties of ball, book, doggie and man, whilst the property of the [Y] slot will be a weighted average of ball, book, rain and pen.
But a weighted average of which properties1: their meanings, their sounds, their stress patterns? In principle, over any of these things; indeed, over any properties that the child can perceive. If the items that appear in a particular position in the source utterances are similar with respect to a given property (e.g., meaning), then the slot in the resultant schema will exhibit this property. So, for the present example, the [X] slot in the a [X] schema will have the semantic property of discrete (“count”) entity, whilst the [Y] slot in the [Y] schema will have the semantic property of discrete or nondiscrete (“mass”) entity. It is important to emphasize that these characterizations are approximate only; the actual meaning of the [X] slot in this example is no more or less than average of the meanings of ball, book, doggie and man; a notion that is captured only roughly by the description “discrete entity”. In other words, slot properties are fuzzy and probabilistic, as opposed to categorical.
If the items that appear in a particular position in the source utterances are dissimilar with respect to a given property (e.g., the sound of the first phoneme), then the slot in the resultant schema will not exhibit any particular property on this dimension. That is, if the source items exhibit heterogeneity with regard to a given property, the slot will also exhibit heterogeneity with regard to this property. So, for example, because the items that give rise to the slot in the the [Y] schema - the ball, the book, the rain and the pen – do not share any particular phonological properties, so the slot does not exhibit any particular phonological properties either (for discussion of the role of variability in slot formation see Bowerman & Choi, 2001; Bybee 1995; Janda, 1990; Barðdal 2009; Suttle & Goldberg, 2011; Dąbrowska & Szczerbinski, 2006).
The significance of slot properties is that only items whose properties overlap sufficiently with those of the slot may be inserted grammatically into this slot (e.g., Langacker, 2000: 17). This notion of overlap is also fuzzy and probabilistic, rather than deterministic. Consider, for example, our example schema a [X], in which the slot has the approximate semantic property of discreteness. Words that exhibit this property to a sufficient degree can be inserted into this slot (e.g., a cat; a table). If a word that does not exhibit this property to a sufficient degree is inserted into this slot, a less than fully grammatical string results (e.g., *a rain). But if we insert a borderline case, something that has an intermediate degree of discreteness (e.g., milk, which is generally continuous, but could denote a discrete serving), an intermediately-grammatical string results (e.g., ?a milk). Consider now our example schema the [Y]. Because this slot has broader semantic properties (“discrete or nondiscrete entry”), we can use pretty much any “entity” noun as a slot-filler (e.g., the milk; the water).
The reason for giving such a detailed account of the acquisition of the English determiner system is that, the account presented above is a microcosm of the constructivist account of language acquisition in general (or, at least, of rote phrases and schematization; a third stage – analogy - is outlined in the section on basic word order). The process by which rote-learned strings give rise to schemas whose slots exhibit probabilistic semantic, phonological and pragmatic properties is assumed to operate in all domains of language acquisition, and across all languages.
Implications of the constructivist account of determiner acquisition
Before moving on to some of these other domains, we consider some broader implications of the account of determiner acquisition outlined above. The first is that, because slots take on whatever properties are shared by the items that appeared in the relevant position in the source utterances, ignoring dimensions along which these items do not share a particular property, there is no need to specify in advance which types of properties children will “look for” when forming grammatical generalizations. This is just as well, since the types of properties that slots exhibit vary hugely cross-linguistically, including – for example – humanness, animacy, and whether or not the speaker witnessed the event. That said, we would not wish to exclude the possibility of very general attentional or perceptual biases that make – say – humans, speech sounds, or the ends of utterances particularly salient.
The second implication is that, because the slot-formation process is sensitive to commonalties along (in principle) any dimension, many slots exhibit constellations of properties of different types. Indeed, to find examples of slots that exhibit semantic, phonological and pragmatic properties at the same time, we need look no further than the English determiner system. Consider the fact that, before nouns that start with a vowel, speakers must use an instead of a. Whilst the traditional approach has been to posit pronunciation variations of “the same” word, this phenomenon falls naturally out of the present account, on the assumption that there are two indefinite constructions – a [X] and an [Z] that have the phonological properties of starting with consonant and vowel sound respectively.
The different pragmatic functions of the and a/an can be accommodated in the same way. The slots in the schemas a [X] / an [Z] and the [Y] have the functional-pragmatic properties of referring to discourse-old and discourse-new entities respectively. Thus, the slot in the schema an [Z] exhibits, at the same time, semantic (discrete entity), phonological (starts with a vowel) and pragmatic (discourse-new) properties. An infelicitous utterance results if the speaker uses a filler in a slot with which it does not share sufficient overlap on any one of these properties (e.g., *an advice [semantic mismatch]; *an cat [phonological mismatch]; *an orange [a pragmatic mismatch, assuming that we have already been talking about this orange). Incidentally, we note in passing that accounts under which children have innate knowledge of a DETERMINER and NOUN category and a rule for combining them (e.g., Valian, 1986; Yang, 2010), will still need to posit something very like this type of probabilistic semantic, phonological and pragmatic learning to account for such cases anyway.
A final, and perhaps controversial, implication of this account is that there is no need for learners to construct either conventional Universal Grammar style grammatical categories such as [NOUN] and [DETERMINER], or even finer-grained language-specific, input-based categories such as [ENGLISH MASS NOUN] or [ENGLISH TRANSITIVE VERB]. The reason is that, provided that the child stores the semantic, phonological (etc.) properties of every word and every construction slot, which she will have to do anyway, she already knows which words can be used in any given construction: the words whose properties overlap sufficiently with those of the construction. There is nothing to be gained by grouping together the words that can act as fillers for the same slot, and calling them a “category”.
Indeed, there are at least two good reasons not to do so. The first is that the compatibility between particular words and constructions is a matter of degree; indeed many words can be coerced into constructions with which they are only marginally compatible (often to yield some special interpretation). The second, and more fundamental, reason for not maintaining offline categories of slot-fillers is that, in many cases, it is impossible to do so. Many properties, such as being discourse-old or discourse-new, which governs the relative compatibility between a noun and the a [X] and the [Y] constructions are not inherent properties of words at all. Hence it would not be possible to store a category of “discourse-old nouns” (or AGENTs or PATIENTs). To the extent that these are “categories” at all, they are ad-hoc categories (Barsalou, 1983), generated on the fly, emergent from the search for a suitable slot filler. If this search can turn up a filler whose pragmatic properties overlap sufficiently with those of the slot, it can presumably turn up a filler whose semantic and phonological properties overlap sufficient with those of the slot just as easily.
3. Inflectional morphology
The constructivist account of the acquisition of inflectional morphology is essentially the account of determiner acquisition outlined above, translated into a new domain. Consider, for example, the acquisition of the German present tense inflectional paradigm for a regular (-en) verb (e.g., spielen, ‘to play’).
1sg Ich spiele (I play) 1pl Wir spielen (We play)
2sg Du spielst (You play) 2pl Ihr spielt (You play)
3sg Er/Sie spielt (He/She plays) 3pl Sie spielen (They play)
The child begins by acquiring a number of rote-learned concrete holophrases, at both the level of individual verb forms and subject+verb combinations:
spielt Er+spielt (He plays)
bekommt Er+bekommt (He gets)
trinkt Er+trinkt (He drinks)
The child then abstracts across these stored strings to yield the following lexically-specific slot-and-frame schemas:
[X]t Er [Y]t
Again, we use [X] and [Y] rather than traditional linguistic category labels (e.g., [VERB] or [STEM]) to emphasize the fact that children are assumed not to be in possession of such categories. Exactly as for the determiner schemas discussed above, these slot-and-frame patterns are productive in that children may insert any filler whose properties overlap sufficiently with those of the slot. Thus, considering semantic properties, children may insert a word denoting an action, event or state-of-affairs (e.g., Er [lieb]t; He likes), but not one that denotes – for example – a concrete object. In this case, these slots do not exhibit any particular phonological properties, though will we subsequently discuss some that do.
As for the determiner system, it is important to emphasize that the constructivist account does not claim that all of children’s early morphological knowledge consists of rote-learned strings. Although this schematization process is long and protracted, it begins as soon as children have, in principle, two stored forms across which to schematize. Thus, on the one hand, even very young children are likely to have acquired slot-and-frame schemas whose source forms are highly frequent in the input (e.g., 3sg, Er [Y]t). On the other hand, even relatively old children are unlikely to have acquired schemas whose source forms are extremely infrequent (e.g., 3pl, Sie [Z]en). Thus the constructivist prediction again relates not to age (“children’s knowledge of morphology is concrete until age X”) but to unevenness: Children will show good performance when they are able to use either (a) a rote-learned ready-inflected form or (b) a slot-and-frame schema formed on the basis of frequent exposure to suitable source utterances, but poor performance when they are not.
There is considerable support for this prediction. For example, highlighting the role of rote-learned ready-inflected forms, a naturalistic-data study of child Spanish (Aguado-Orea, 2004) found that an error rate of 5% for 1sg forms doubled to 10% when just the two most frequent – and hence potentially rote-learned – forms (quiero, ‘I want’ and puedo, ‘I can’) were excluded. This study also found that both children studied used significantly fewer different forms of each verb than their caregivers. Importantly, this finding is not simply a consequence of adults having a larger repertoire of verbs or morphemes; it holds even when restricting the analysis to verbs and morphemes used by both children and their caregivers. Krajewski (2008) reported very similar findings for Polish (though similar studies of English, Spanish and Italian – Pine, Lieven & Rowland, 1998; Gathercole, Sebastian & Soto, 1999; Pizzuto & Caselli – failed to include these crucial vocabulary controls).
Considering the role of morphological slot-and-frame schemas, Aguado-Orea (2004) found an error rate of close to zero for 3sg forms, but 34-58% (depending on the child) for 3pl forms; exactly the finding that is expected, given that source forms for a 3sg [X]a schema are considerably more frequent in the input that source forms for a 3pl [X]an schema. Similar findings were reported in a naturalistic data study of Brazilian Portuguese (Rubino & Pine, 1998), and an elicited production study of Finnish (Rasanen, Ambridge & Pine, submitted).
Although its impoverished morphology makes translating such studies into English less than straightforward, when this is done, similar findings are obtained (e.g., Theakston, Lieven & Tomsello, 2003; Theakston & Lieven, 2005; Theakston, Lieven, Pine & Rowland, 2005; Theakston & Rowland, 2009; Rowland & Theakston, 2009; Wilson, 2003; Pine, Conti-Ramsden, Joseph, Lieven & Serratrice, 2008; Rasenen, Ambridge & Pine, in press). Neither are such findings restricted to verb morphology. Similar findings for noun morphology were reported by Leonard, Caselli and Devescovi (2002) for Italian, and by Dąbrowska (2001; 2004; 2005; 2008a; 2008b), Dąbrowska & Szczerbinski (2006) and Krajewski, Lieven and Theakston (2012) for Polish (though again, there is some evidence of the beginnings of abstraction from an early age; Dąbrowska & Tomasello, 2008; Krajewski, Theakston, Lieven and Tomasello, 2011).
The role of phonology
The morphological systems discussed in this section so far do not have any particular phonological restrictions. Many systems, however, do exhibit such restrictions, two particularly well-studied examples being the Arabic and German noun-plural systems (Forrester & Plunkett, 1994; Hare, Elman & Daugherty; Plunkett & Nakisa, 1997; Kopcke, 1998; Hahn & Nakisa, 2000; Kauschke, Kurth & Domahs, 2011; Behrens, 2002, 2011). Such restrictions are incorporated into the present framework by an assumption already introduced above; that slots exhibit phonological properties which derive from a weighted average of the items that appeared in this position in the source utterances. For example, the German –s plural marker appears with a wide variety of phonologically heterogeneous nouns. Consequently the [X] slot in the morphological [X]s slot-and-frame schema has no particular phonological properties, and can readily accommodate all-comers (including, for example, foreign borrowings with non-nativelike phonology). Other markers, such as –e appear with a smaller set of phonologically homogenous nouns (the vast majority end in t/d, with a much smaller number ending in –s/-r/-k). Consequently, the [Y] slot in the morphological [Y]e slot-and-frame pattern probabilistically exhibits the phonological property of ending in –t > -d > -s/-r/-k, and the greater the extent to which a novel noun is compatible with this property, the more likely it is to receive the –e plural marker, as opposed to one of the alternatives (e.g., Kopcke, 1998; Hahn & Nakisa, 2000). The same holds for any other dimensions along which slot-fillers share a particular property (e.g., gender and animacy are two such dimensions for some German noun classes).
Type frequency versus heterogeneity
The account outlined above raises an interesting question: Is a particular slot more productive (more open to new fillers) when it has higher “type frequency” - i.e., when more different items have appeared in this position in the source utterances (Bybee, 1995; 2001) – or is type frequency simply functioning as a proxy for heterogeneity? Certainly, all other things being equal, the greater the number of unique items that appear in a particular position, the greater the likelihood that these items will be heterogeneous (in terms of phonology, semantics, gender etc.), thus giving rise to a slot that exhibits no particular properties, and hence no particular restrictions. Findings from investigations into the German noun plural suggest the latter possibility: the [X]s schema appears to be one of the most productive schemas (see also Behrens, 2011, on –en), although it has considerably lower type frequency than many others, presumably due to its diversity.
Although the findings from an adult study (Suttle & Goldberg, 2011) suggest unique roles for type frequency and (here, semantic) heterogeneity, we are aware of only one child study that has attempted to unpack these two factors. In a novel-noun elicited-production study of the Polish case-marking system, Dąbrowska and Szczerbinski (2006) found that phonological heterogeneity was a better predictor of productivity than type frequency for 4-year-olds and adults, whilst the opposite was true for 2-year-olds. For three-year-olds, the two predictors were similar. However, these results should be interpreted with caution, given the extremely high correlation between these authors’ measures of type frequency and phonological heterogeneity (r=0.91, p<0.001). Future research should focus on morphological systems that better allow these predictors to be unconfounded.
The effect of phonological neighbourhood density
In the meantime, a related finding that warrants explanation is the apparent importance of phonological analogy, even for slots that do not place phonological restrictions on the fillers that may appear therein. Consider, for example, the Finnish present-tense 1sg morpheme –n. Due to the highly regular nature of Finnish morphology, all verbs, regardless of their phonology, take –n in 1sg form. Why, then, are children more likely to supply the correct form for verbs with a large number of phonological neighbours (kerää-n ‘I pick up’, herää-n ‘I wake up’) than for verbs that do not (Rasanen, Ambridge & Pine, in press; see also Kirjavainen, Nikolaev & Kidd, 2012, for a similar finding for the Finnish past-tense system)? Why, in a similar vein, are English children more likely to supply a regular –t (orthographically, -ed) past-tense form for real and novel verbs that are similar to several existing regular verbs (kiss/kissed, miss/missed, hiss/hissed, wish/wished) than those that are not (e.g., match) even though all are compatible with the [X]t schema (Marchman, 1997; Marchman, Wulfeck, & Weismer, 1999; Ambridge, 2010)?
One possible explanation for these phonological neighbourhood density effects is that, rather than using slot-and-frame schemas, children sometimes arrive at an inflected form by direct phonological analogy with one or more stored rote-learned, ready-inflected forms. However, this explanation seems a little ad hoc; why should children rely on one-shot analogies more in the domain of morphology than elsewhere? An explanation that is more consistent with the general approach that we are outlining here emphasizes that the property of a slot is a weighted average of the properties of all the items that have appeared in the relevant position in the source utterances. Thus even when these items are sufficiently heterogeneous as to mean that a slot exhibits no restrictive phonological properties, it nevertheless retains a preference for fillers that share the “flavor” of its source items, in the form of a weighted average (much as we propose for sentence-level categories in the following section). For example, to put some crude numbers on it, the [X] slot in the English past-tense [X]t schema might have the phonological properties 30% iss (kiss, miss, hiss), 20% ish (e.g., wish, dish), 50% miscellaneous (e.g., clip, bake etc.); sufficiently variable that any novel verb can be accommodated, sufficiently weighted that a novel verb such as wiss or biss feels most at home (i.e., is rated as a particularly “good sounding” regular past-tense form; see Albright & Hayes, 2003; Ambridge, 2010).
There is also a third, and perhaps most satisfactory, possibility2; one that we have already touched upon in the section on determiner acquisition. Perhaps what constructivist theorists call schemas or constructions are no more than helpful mnemonics for particularly frequent kinds of generalizations. Perhaps, in fact, all generalizations are formed on the fly on the basis of stored strings, with schemas merely “immanent in their instantiations” (Langacker, 2000: 7). On this account, there is no difference between generating wisswissed (a) by phonological analogy with kissed, missed and hissed or (b) by inserting wiss into a [X]t slot-and-frame schema, with which – by virtue of the occurrence of kiss, miss and hiss in its source utterances – it is particularly compatible. In fact, although rarely discussed in relation to child language acquisition (notable exceptions are Ninio, 1993 and Abbot-Smith & Tomasello, 20063), such exemplar-based models of linguistic knowledge are well established in the adult psycholinguistics literature (e.g., Skousen, 1992; Nosofsky, 1992; Chandler, 2002; Daelemans, Zavrel, Van der Sloot & Van den Bosch, 2010).