We join other responders in thanking Clahsen and Felser (CF) for pulling together these observations about the development of language processing. We are especially impressed by the generality and inclusiveness of CF's treatment of development in L1 and L2. Because most of their specifics concerned comprehension processes, our contribution will emphasize the added value of an appeal to production processes. In particular, we will articulate the value of applying existing production models to developmental phenomena. Language development can be interpreted in revealing ways through the lens of adult sentence production models. These models specify how lexical, syntactic, morphological, and phonological knowledge are integrated in real time as we produce sentences. They are performance models, but they go well beyond general measures of working memory and general notions of limited capacity and resources.
PRODUCTION MODEL PERSPECTIVES ON DEVELOPMENTAL PROCESSES
In the past 40 years, research in adult language production and in child language acquisition has evolved along divergent paths despite being subdisciplines of psycholinguistics. Research on adult production uses systematic error to deduce the architecture and timing of processes like message formation, lexical access, sentence planning, and phonological encoding. This research has given rise to elaborate and revealing models of real-time sentence production (e.g., Bock & Levelt, 1994; Dell, 1986; Fromkin, 1971; Garrett, 1984). In contrast, systematic error in children's utterances is usually assumed to reflect limitations in their linguistic competence (e.g., Radford, 1990; Rice, Wexler, Marquis, & Herschberger, 2000; Wexler, 1996). Essentially, research on adult production has studied performance and real-time processes, whereas research on child production has studied linguistic competence.
Distinguishing between competence and performance in children seems more challenging than it is in adults, in part for methodological reasons. This difficulty may have led some to avoid children's performance limitations as an essential aspect of development or even to think that such limitations are not potentially informative or worthy of research (e.g., Crain & Thornton, 1998). CF clearly show the value of studying the development of parsing in child and adult language learners. As they note, little research addresses these issues. Even fewer studies concern the developing production system. This is striking, given the large body of production data in the field. Although obviously relevant, it is rarely used to explore the production system itself. Wijnen (1990, p. 651), echoing Marshall (1979), made this point 14 years ago: “Although most child language research is based on spontaneously produced speech, the predominant approach is competence-oriented.” The irony here is that production data are arguably poorly suited to studying adult grammar (for all the reasons that account for its motivated exclusion by syntacticians working on adult language). Both the adult and child language data represent competence through a production filter. By and large, we know how to exclude adult error from our data base. For early child language, the matter of “error” is intrinsic, so a satisfactory analysis must provide principled ways to evaluate the effects of that filter. For such analysis, we need a theory of the real-time production system.
The few existing studies of the developing production system emphasize phonological and lexical speech errors in spontaneous utterances (e.g., bolar pears for polar bears; easy for hard; Stemberger, 1989; see also Jaeger, 1992; Wijnen, 1992). Children's speech errors are revealing. Indeed these small n studies show that some aspects of adultlike speech errors are present in early childhood. Interestingly, these aspects appear developmental in nature. Still, Bock (1991, p. 156) points out that “the linguistic and extralinguistic contexts of natural errors may vary freely in ways that bear both on the occurrence of error and on the validity of different explanations, making some form of experimental control essential.” However, the experimental paradigms and large n studies that are needed to understand the developing system are largely lacking.
Some developmental psycholinguists, including CF, do consider children as both acquiring new knowledge and learning to use this knowledge in the split-second operations of sentence production. We would like to note a few additional studies in this vein. Some components of sentence production, such as self-monitoring, clearly develop (Rispoli, 2003). The vocabulary spurt observed in young children may be due to a restructuring of the lexical retrieval part of the production process rather than to an actual increase in vocabulary size (Dapretto & Bjork, 2000). Young children appear to have capacity limits that may initially restrict the length and complexity of their utterances (Rispoli & Hadley, 2001). The classic phenomenon of subject omission may also involve capacity limitations (Bloom, 1990; Valian, Hoeffner, & Aubrey, 1996). Morphosyntactic elements that are inserted during later stages of sentence production may initially be difficult to generate (McKee & Iwasaki, 2001). Cross-linguistic variation in omission patterns suggests that the way a language packages lexical information also affects these processes (McKee & Emiliani, 1992). Children's difficulties with some relative clause constructions may be due to the developing formulator's adherence to strategies that aid the sentence planning process rather than to strategies that aid the parser (McDaniel & Lech, 2003). Children's pauses in complex sentences occur in the same positions as those of adults, indicating a similar planning process (McDaniel, McKee, & Garrett, 2005). Such studies are just the beginning. Much more warrants study.
CF discuss at some length the production study by Clahsen, Hadler, and Weyerts (2004), which addressed the linkage of knowledge and processing during development. Clahsen et al. (2004) found that, although children's participle forms were essentially the same as in adult's production, children were generally slower. Children also showed an antifrequency effect in response latency for regular participles. That is, high-frequency regular participles took longer to produce than low-frequency regular participles. This effect disappeared in adulthood. Pinker (1999) suggests an explanation for these findings via the competition between two mechanisms for the production of morphologically complex forms, one based on the retrieval of a word form and the other on the use of a morphological rule. This competition appears most prominently for forms in which high-frequency representations of regularly inflected words are stored independently of the rule route and can inhibit the application of the rule route.
Clahsen et al. (2004) is a case that exemplifies the need to “developmentalize” production models. Now let us situate affix production into a particular model of real-time sentence production. This history of the model begins in the 1970s (Fromkin, 1971; Garrett, 1975) and has evolved over a 30-year period (Bock & Levelt, 1994; Dell, 1986; Garrett, 1984; Levelt, 1989). In this model, sentence production and sentence comprehension are codependent but separate systems. Of central importance is a division in the word processing system between lemma level semantic and syntactic information and lexeme level phonological information. The model is, in important respects, lemma driven. This means that the information contained in the lemma for a word mandates syntactic frames.
Critical features of the model are motivated by speech errors, including lexical exchanges. Lexical exchange errors may strand affixes, and this reveals important details of the production process. For example, a speaker who intends to say “bless Dr. Schwartz's heart” may say instead “bless Dr. Heart's Schwartz” (Pamela Hadley, personal communication). This observation motivates the hypothesis that inflectional affixes are part of the grammatical encoding, and not inserted lexically. That stranded affixes typically adjust to the environment of the incorrectly inserted lexical stem buttresses this hypothesis (Garrett, 1975). Such errors show that affix production has two stages. The early stage uses the syntactic/semantic information in the lemma, whereas the later stage uses the phonological information in the lexeme (Levelt, 1989). This model also posits a process called indirect election, by which lemma information mandates the incorporation of inflectional affixes into a syntactic frame (Bock & Levelt, 1994).
As a concrete example involving affixation, we can now ask how children's antifrequency effect of regular forms might play out in a model of sentence production (Clahsen et al., 2004). It is likely that some lemmas, even for high-frequency regular words, are incomplete during development. Whole word forms for both irregulars and regulars might initially be stored unanalyzed. Those representations would compete with a rule-driven system as the latter develops (in our terms, by the elaboration of a lemma system that either calls up the appropriate affixation process for regular forms or blocks it for irregular forms). The competition between whole form representations and rules must be maintained for the irregular forms. The entries in that system must continue to inhibit the rule-based system. If high-frequency items are initially stored as unanalyzed wholes and that set includes both irregulars and regulars, then the regular high-frequency forms will (counterproductively) inhibit the rule-based system. Which lemmas will have to undergo restructuring must somehow get sorted out: the learner must eventually know which whole form entries should win the competition (the irregulars) and which can run by either output route (the regulars). Although incomplete in many details, this story about a decision conflict could explain the (surprising) antifrequency effect. (Note, though, that the architecture described above accommodates but is not necessary to this explanation.)
Theories of adult sentence production were designed to account for adult language use. Although they have been usefully applied to the study of language breakdown (e.g., aphasia, Alzheimer disease), they are rarely applied to learners who are building up both language knowledge and speaking proficiency. The problem we see is that the models generally do not articulate developmental hypotheses. Do lemma-driven models of (adult) sentence production predict what children do, if indeed the lemma must be taken into account to understand development? What does the proceduralization of morphosyntax imply for lexical storage? Adult sentence production models have evolved by encompassing a range of phenomena exhibited in aphasic syndromes (Foygel & Dell, 2000; Garrett, 1982; Lapointe, 1985). Application of these models to aphasia has tested the models and necessitated their revision. In other words, these models have been seriously advanced by the study of aphasia. Developmentalizing them should have similar positive effects. Indeed, one might argue that accounting for language development is even more important than accounting for language breakdown, because the production system must develop, but it may or may not break down later on.
Sentence production models
Here are some research issues for which model-theoretical details might be informed by developmental considerations. Models of language production have developed along similar lines in some areas but show sharp contrasts in others. Future research should find the leverage points in production modeling that can be convincingly tied to developmental data and find ways to assimilate existing developmental claims with one or another approach to production modeling.
Some obvious areas for exploration involve the diverse relations between conceptual structure, lexical representations, and sentence integration processes. For example, some models of lexical retrieval map directly from the conceptual level to word form (e.g., Caramazza, 1997), while most others assume an indirect mapping that links abstract word representations with conceptual structure and sharply separates this from phonological and articulatory stages of representation and process. Within this latter group, there are differences in the extent to which interactive processes hold between semantic, syntactic, and phonological representation (e.g., Dell, 1986; Dell, Schwartz, Martin, Saffran, & Gagnon, 1997; Ferreira, 2003; Levelt, Roelofs, & Meyer, 1999; Vousden, Brown, & Harley, 2000). Closely related questions arise in debates about the relation between conceptual representation and aspects of morphosyntax such as number agreement, gender marking, and pronominal expression (e.g., Bock, Eberhard, & Cutting, 2004; Bock & Miller, 1991; Vigliocco & Hartsuiker, 2002). More general issues of conceptual effects on lexical retrieval for sentence integration have been explored in several works by Bock and colleagues (e.g., Bock, 1987; Bock, Loebell, & Morey, 1992). How, precisely, do the competing adult accounts accommodate child language based claims for relations among lexical, syntactic, and semantic processes?
Another area that warrants exploration involves relations between comprehension and production systems. Most production models assume systems with distinct comprehension and production mechanisms, along the lines sketched by Levelt (1989). However, variation in the interaction of the systems is played out in theories of production monitoring (e.g., Postma, 2000). This leads to different views about the extent of overlap in the lexical and syntactic systems that serve production and comprehension. One of the strongest claims in this area is Kempen's (2004) proposal that the same mental machinery is responsible for syntactic integration in both production and comprehension. How, precisely, do the competing adult models accommodate the relative emergence of production and comprehension capacities in child language?
Developmental phenomena and performance-based explanations
Here are some complementary ways that patterns in children's phonology, lexicon, syntax, and discourse might be informed by consideration of production models. Again, future research should tie the production models to developmental data and see how developmental phenomena affect one or another approach to production modeling.
Some areas for exploration involve continuity questions. For example, does the architecture of the production system change radically over time or can changes in children's production patterns be explained by increasing processing resources alone? Most of the research that CF review suggests that increasing resources affects comprehension. In a case study, Wijnen (1990) found that a child's dysfluencies were initially randomly distributed but later concentrated on function words and sentence-initial words. This may indicate that as learners master a grammar, their errors reflect the planning frames supported by that grammar. Such an interpretation is suggested by the child's dysfluencies across structures: some structures hosted more dysfluencies than others.
Such patterns suggest that an examination of frequency issues would also be useful. Wijnen (1992) reported a higher overall frequency of speech errors in children compared to adults but a lower incidence of phonologically similar lexical substitutions. He also found that children, un like adults, make sound errors in the high frequency functional elements. Children's production of certain syntactic structures can also be studied from this perspective. Learners' utterances reflect frequency patterns in their input, patterns that are less clear in their grammaticality judgments (McDaniel & Lech, 2003; McKee & McDaniel, 1998). We might ask to what extent frequency-based representations characterize the competence grammar and whether those representations are fundamentally the same (or continuous) over the course of language development.
Relations between comprehension and production systems are of great interest in child language, and the potential links to adult processing systems are intriguing. It is a truism that children's comprehension capacity outstrips the structure shown in their spontaneous production. The apparent dissociation is robust, but theoretical accounts of it are not. Moreover, the empirically supported range of structural contrasts between production and comprehension in child language is still rather limited. Thus, this is an area that also needs exploration. Explaining such differences provides a clear challenge for assumptions about relations between processing models and competence models in child language.
