This article discusses issues in the linguistic analysis of morphosyntax and syntax in children with language disorders. It presupposes familiarity with generative linguistic theory. The goal is to illustrate what kinds of grammatically based explanations are available to theories of disorders. The task of such explanations is to unify superficially diverse markers of a disorder into natural classes that linguistic theory motivates or to characterize observed errors. The discussion is divided, as suggested by recent generative theory, into issues concerning the lexicon versus issues concerning the computation of larger structures from lexical atoms.

The lexicon can be subdivided along various dimensions, any of which might be relevant in capturing dissociations found in nonnormal language development. A primary split divides content words, i.e., nouns, verbs, adjectives, and adverbs, from all other morphemes, i.e., functional or closed-class elements. (I ignore the challenging issue of where adpositions belong; they likely are heterogeneous, a dichotomy among them being evinced in adult disorders [Rizzi, 1985; Grodzinsky, 1990].) In Chomskyan syntax of the 1980s (cf. Emonds, 1985; Chomsky, 1986; Fukui, 1986) there was a seemingly arbitrary split between functional meanings that were treated as autonomous syntactic functional categories (e.g., Tense, Determiner, Agreement, Complementizer) and those that were not (e.g., number marking on nouns, participial affixes, infinitival suffixes on verbs). Some attempts were made to understand disorders of acquisition in terms of this heterogeneous system (cf. Leonard, 1998). However, the harder linguists have looked, the more functional heads for which they have found structural evidence. On parsimony grounds we now expect all functional meanings to be represented in syntactic positions separate from those of content words (cf. van Gelderen, 1993; Hoekstra, 2000; Jakubowicz and Nash, 2001).

Among functional elements (morphemes or features thereof) a further distinction is made, dubbed “interpretability” by Chomsky (1995). Some morphemes encode elements of meaning and contribute directly to the interpretation of a sentence; for example, the English noun plural suffix -s combines with a noun whose meaning describes a kind of entity, and adds information about number; for example, dog + s = caninehood + more-than-one. Plural -s is therefore an interpretable morpheme. This contrasts with the 3sg present indicative verbal suffix -s, which makes no semantic contribution, because what is semantically “one” or “more than one” is the subject of the sentence—a noun phrase. Numerosity (and person) are properties of noun meanings, not verb meanings. The inflection -s appears on the verb as a consequence of the number and person of the subject, that information having been copied onto the verb. Agreement is therefore an uninterpretable morpheme: it replicates a bit of meaning represented elsewhere, surfacing only because the morphosyntax of English requires it. It has no impact at Logical Form. Such morphemes are seen as a plausible locus for impairment (cf. Clahsen, Bartke, and Göllner, 1997).

A second common kind of uninterpretable morphology is case marking. Many case markings have nothing to do with meaning. For example, in She saw me and I was seen by her, the difference between accusative me and nominative I does not correspond to any change in the semantic role played by the speaker; rather, it reflects purely syntactic information. A third kind of uninterpretable morphology arises in concord. For instance, in Latin the forms of many words within a noun phrase reflect its case and number, as well as the gender of the head noun: for example, ill-as vi-as angust-as “those-acc.fem.pl streets-acc.fem.pl narrow-acc.fem.pl.”

Among the uninterpretable features there may be a further distinction to be drawn, as follows. Person and number information is interpretable on noun phrases but not on verbs; in a sense, there is an asymmetry between the contentful versus the merely duplicated instantiation of those features (cf. Clahsen, 1989). Case is different: it is taken to be uninterpretable both on the recipient/checkee (noun phrase) and on the source/ checker (verb, preposition, or Tense); it does not exist at Logical Form at all. Thus, symmetrically uninterpretable functional elements such as Case constitute another natural class. Grammatical gender is also an example of a morphosyntactic feature that has no semantic counterpart. There are two ways in which gender marking might be impaired, calling for different explanations: first, children might not be able to consistently recall the gender of particular nouns; second, children might not be able to consistently copy gender information from the noun, which is associated with gender in the lexicon, to other parts of a noun phrase (a problem with concord).

Turning now from the lexicon to the syntax, under Minimalism the transformational operations Merge and Move are restricted in a way that effectively builds some former filters into their definitions. Taking this seriously, certain ways of talking about language disorders in terms of missing or dysfunctional subcomponents of syntax no longer make much sense. For example, in the theory of Chomsky (1981), it was plausible to talk about, say, bounding theory (which encompasses the constraints on movement operations) being inoperative as a consequence of some disorder; this simply meant that certain structures that were always generable by Move a were no longer declared invalid after the fact by virtue of a movement that was too long. There is no natural translation of this idea into Minimalist machinery because locality is part of what defines an operation as a Move. A language disorder could in principle involve a change in this definition, but we cannot think of this as simply excising a piece of the grammar. Similarly, since movement in Minimalism is just a means to an end, it makes little sense to think of eliminating movement while leaving the rest of the theory intact; the system is now designed such that movement is the only way to satisfy a fundamental requirement that drives the computational system, namely, the need to create valid Logical Forms by eliminating uninterpretable features.

The syntactic accounts of language disorders that have been suggested often involve impoverished structure. One way to execute this is to posit that particular functional heads are either missing from syntactic structures or do not contain (all) the features that they would for adults (cf. Wexler, 1998). Another way is by reference to the position of the heads in the clausal structure rather than by reference to their content. Thus, it can be proposed that all structure above a particular point, say VP, is either missing (Radford, 1990) or optionally truncated (Rizzi, 1994). The extent to which particularly the tree pruning variant of this idea (Friedmann and Grodzinsky, 1997) can be characterized as simply lesioning one independently motivated piece of grammar is open to debate.

The division between stored representations and computational procedures is also relevant in (particularly inflectional) morphology, where the contrast between general rules and stored probabilistic associations has a long history (Pinker, 1999). Where precisely the line should be drawn differs, depending on one's theory of morphology. For instance, if there are any rules at all, then surely the process that adds (a predictable allomorph of) -d to form the past tense of a verb is a rule, and unless all morphology is seen as executed by rules, the relationship between am and was is stored in the lexicon as an idiosyncratic fact about the verb be, not encoded as a rule. But in between lie numerous subregularities that could be treated either way. For example, the alternation in sing-sang, drink-drank, and so on could be represented as a family of memorized associations or as a rule (with some memorized exceptions) triggered by the phonological shape of the stem ([ıŋ]). Most psychologically oriented research has assumed that there is only one true rule for each inflectional feature such as [past tense] (supplying the default), while the generative phonology tradition has used numerous rules to capture subpatterns within a paradigm, only the most general of which corresponds to the default in having no explicit restriction on its domain. Thus, although a dissociation between rules and stored forms is expected under virtually any approach, precise predictions vary.

There are several ways in which (inflectional) morphology might be impaired. One of the two mechanisms might be entirely inoperative, in which case either every inflected form would have to be memorized by rote (no rules; cf. Gopnik, 1994) or every word would have to be treated as regular (no associations). The latter would be evinced by overregularizations. The former might yield overirregularizations if, in the absence of a rule, analogy-based associations were unchecked and overextended; alternatively, it might have a subtler symptomology whereby any verb could be correctly learned once its past tense was heard, but in a Wug-testing situation nonce forms could not be inflected. A more moderate impairment might entail that at least some irregular forms would be demonstrably learnable, but in production they would not always be retrieved reliably or quickly enough to block a general rule (Clahsen and Temple, 2002), apparently violating the Elsewhere Condition.

We turn now to the logic governing ways in which the most commonly produced error types from the child language disorder literature can be explained. Under the strong separation of computational combinatoric machinery versus lexical storage pursued in Minimalist syntax, it is possible for children with normal syntactic structures to sound very unlike adults, because in their lexicon certain morphemes either are missing or have incorrect features associated with them. For example, the fact that some children learning English might never produce -s in an obligatory 3sg context could be consistent with them having mastered the syntax of agreement, if their lexicon has a missing or incorrect entry for 3sg. Consequently, it is important to document the lexical inventory, that is, whether each child does at least sometimes produce the forms of interest or can in some way be shown to know them. Similarly, suppose that children learning a given language produce agreement mismatch errors of just one type, namely, that non-3sg subjects appear with 3sg verb forms. That is, suppose that in Spanish we find errors like yo tiene (“I has (3sg)”) but no errors like ella tengo (“she have (1sg)”). One could postulate that 3sg forms (e.g., tiene) are unspecified for person and number features and are inserted whenever more specific finite forms (e.g., tengo) are unavailable, or that tiene has wrongly been learned as a 1sg form. In either case, again, the syntax might be fully intact.

Errors due to lexical gaps should pattern differently from errors due to syntactically absent heads. Consider the Tense head as an example. The meaning of tense itself (past versus present) can be expressed morphologically in the Tense head position, and in addition Tense is commonly thought to house the uninterpretable feature that requires the presence of an overt subject in non-null subject languages (the unhelpfully named EPP feature), and the uninterpretable feature that licenses nominative case on its specifier (the subject). If Tense were completely missing from the grammar of some children, this would predict not only that they would produce no tense morphemes, but also that they would not enforce the overt subject requirement and would not (syntactically) require nominative case on subjects. If, on the other hand, what is observed is not an absence of tense marking but rather an incorrect choice of how to express Tense morphophonologically (e.g., singed, instead of sang), this would not be compatible with absence of the Tense head and would not predict the other syntactic consequences mentioned. Only omission of an inflectional marking or perhaps supernumerary inflection (e.g., Did he cried?) could have a syntactic cause; incorrect allomorph selection could not.

If a feature such as Tense is expressed some of the time but not always, this could have two underlying causes that lead to different syntactic expectations. One possibility is that Tense is part of the syntactic representation in all cases and its inconsistent expression reflects some intermittent problem in its morphophonological spell-out (cf. Phillips, 1995; Marcus, 1995); in that case no syntactic consequences are predicted. The other possibility is that Tense is intermittently absent from syntactic representations. In just this scenario we expect that the syntactic properties controlled by the Tense head should be variable (e.g., subjects are sometimes nominative, sometimes not), and furthermore, utterance by utterance, syntax and morphology should correlate: the Tense morpheme should be missing if and only if nominative case is not assigned/checked and the overt subject requirement is not enforced.

It is crucial to understand such predicted contingencies as claims concerning the distribution of contrasting forms. It seems clear that some children go through a stage during which, for example, some of the English pronoun forms are not produced at all; this is particularly common for she. What behavior we should expect at this stage depends on assumptions about the architecture of the syntax-morphology interface. Taking an “early insertion” view, under which only complete words from the lexicon can be inserted in syntactic derivations, a child lacking a lexical entry for the features [pron, 3sg, fem, NOM] should be unable to generate a sentence requiring such an item. In contrast, on a “late insertion” nonblocking approach such as Distributed Morphology (Halle and Marantz, 1993), there is always some form (perhaps phonologically null) that can be used to realize the output of a syntactically valid derivation. The syntactic tree is built up using feature bundles such as [pron, 3sg, fem, NOM], without regard to which vocabulary item might fill such a position. The architecture dictates that if there is no vocabulary entry with exactly this set of features, then the item containing the greatest subset of these features will be inserted. Thus, a child who knows the word her and knows that it is a feminine singular pronoun could insert it in such a tree, producing Her goes from a fully adultlike finite clause structure.

We conclude with some methodological points. Generative grammar's conception of linguistic knowledge as a cognitive (hence also neural) representation in the mind of an individual at some particular time dictates that, in analyzing behavioral data collected over an extended period or from multiple children, pooled data cannot be directly interpreted at face value. For example, suppose we analyze transcripts of three children's spontaneous productions sampled over a period of 6 months and find that in obligatory contexts for some grammatical morpheme, say 3sg -s, the overall rate of use is 67%. Virtually nothing can be concluded from this datum. It could represent (among other possibilities) a scenario in which two children were producing -s at 100% (i.e., talking like adults in this regard) and the third child never produced -s. In this circumstance we would have no evidence of a developmental stage at which -s is optional, and the third child's productions would be consistent with her simply not knowing the form of the 3sg present tense verbal inflection in English. At another extreme, if each of the three children were producing -s in two-thirds of obligatory contexts, we would have to posit a grammar in which Tense or Agreement are optional (or else multiple concurrent grammars). The same logic applies in the temporal dimension: a period of 0% production followed by a period of 100% production calls for a different analysis from a period during which production is at 50% within single recording sessions. Therefore, data reporting needs to facilitate assessment of the extent to which samples being pooled represent qualitatively comparable grammars. In addition to measures of central tendency and variance, this calls for the kind of distributional information found in “box-and-whiskers” plots.

Furthermore, to say that a child “has acquired X” by a certain age, where X is some morpheme or class of morphemes, is not strictly meaningful from the linguistic perspective adopted here. We can speak of attaining levels of normal or adult performance, and we can speak of having established a lexical entry with the correct feature specifications, but there is nothing in grammar that could correspond to a claim such as the following: “A child is taken to have acquired X once her rate of production of X morphemes in obligatory contexts is greater than 90%” (cf. Brown, 1973). This begs the question of what the grammar was like when production of X was at 85%: If X was not at that time a part of her grammar, how did the child manage to create the illusion of using it correctly so much of the time? Also, rates of production in obligatory contexts must be complemented by correct usage rates; one is scarcely interpretable without the other. If X is always used correctly, then even very low production rates signal knowledge of the properties of X and the syntactic conditions on its distribution. But if X is also frequently used incorrectly, neither type of knowledge can be inferred.