Theoretical Perspectives on Plural Morphology

One of the holy grails of cognitive science is the explanation of linguistic productivity: a model of the language system that allows people to store what they hear and creatively reuse it to novel ends. There has been extensive debate concerning how one should carve apart storage and combinatorial use, with some who have favored a strict division between the concrete representation of lexical items and their combination in terms of abstract categories (e.g., Chomsky, 1965) and others who have argued that storage is more pervasive and associative in nature from which abstract combinatorial schemas may first arise (e.g., Langacker, 1987). Inflectional morphology has long been the favored test case for these contrasting views of language. The intense research in this field offers a particularly rich set of explanatory alternatives that we explore here from a developmental perspective.

Research into the acquisition of inflectional morphology has tended to concentrate on the phenomenon of overregularization—the application of a regular inflection to an irregular stem or inflected form, for example, mouses or mices. Depending on one’s point of view, such errors of commission have been taken as the quintessential demonstration that children either acquire abstract rules or generalize schemas. Far less attention has been paid to the errors that occur when children simply do not inflect a word at all—for example, saying “mouse” to refer to five mice. Yet these errors of omission can shed considerable light on the acquisition process and can potentially pull apart differing theoretical accounts of inflectional morphology.

The dual-route model of inflection (Clahsen, 1999; Clahsen, Aveledo, & Roca, 2002; Clahsen, Rothweiler, & Woest, 1999; Marcus, 1995a, 1995b; Marcus et al., 1992; Pinker, 1999; Pinker & Prince, 1988; Pinker & Ullman, 2002) proposes that irregularly inflected words (e.g., mice) are stored in an associative memory, whereas the inflection of regular words (e.g., houses) is computed by a default rule (e.g., “add -s” for English plurals) that combines a symbol for a stem with a symbol for a suffix. Before the default rule is acquired, if a child does not have an appropriate inflected form in memory, then she or he will be forced to utter a bare stem in its place (Pinker, 1999), and therefore, errors of omission may occur prior to the acquisition of a rule. On learning the rule (as evidenced by the production of overregularizations), errors of omission should disappear, and the child should inflect consistently, albeit with overregularized forms for unknown irregulars (e.g., mouses). However, the dual-route model allows for analogies to be made between irregular nouns/verbs that form phonological subgroups, allowing a certain degree of erroneous productivity (e.g., bring-brang by analogy to sing-sang, ring-rang). Thus, any subsequent errors of omission could only be explained if one were to argue that the word in question was mistaken to be a “no change irregular” (e.g., cut-cut) by virtue of being a close phonological neighbor of such irregulars. When this is not the case, errors of omission are not predicted.

Usage-based or schema models (Bybee, 1985, 2001; Bybee & Slobin, 1982; Dabrowska, 2001, 2004; Köpcke, 1998) propose that both regulars and irregulars are handled by the same storage and processing mechanisms. In a sense, schema models are also dual-route models in that a given inflected form may be arrived at either by retrieving the whole form from memory or by accessing a stem and adding appropriate affixes, which are represented as schemas (Bybee, 2001). The difference is that the choice of route is determined by the frequency of the form, not its regularity. The higher the frequency of the inflected form, the more likely the whole form is retrieved from memory (whether it is regular or irregular). Bybee and Slobin (1982) proposed that children form product-oriented schemas, which are generalizations about properties of inflected forms (e.g., past tenses tend to end with –ed ) as well as source-oriented schemas, which are generalizations about how an inflected form is composed of a stem and an inflection (e.g., to make a past tense, take a verb stem and add –ed ). Adult-like mastery is achieved by balancing product-oriented schemas with source-oriented generalizations about suffixation and irregular idiosyncrasies. The early development of product-oriented schemas predicts that children will make more errors of omission with nouns and verbs whose stem ending already resembles the inflected schema (e.g., dress already ends in –s). This account does not predict that errors of omission should stop once errors of commission (productive suffixation) begin. Rather, the developmental process is seen as a gradually refined balancing act that is sensitive to the token frequency of the individual words being inflected (be they regular or irregular) and the type frequency of the inflection(s) being schematized.

Connectionist models have generally been compatible with schema models in that they propose the same storage and processing mechanisms for regulars and irregulars (e.g., Joanisse & Seidenberg, 1999; MacWhinney & Leinbach, 1991; Plunkett & Juola, 1999; Plunkett & Marchman, 1991, 1993, 1996; Rumelhart & McClelland, 1986; Westermann, 1998). These models show signs of schema-induced errors of omission (e.g., making more errors of omission with the /-id/ inflection required for verbs whose stem already ends in d/t), although these errors are not very frequent and are often argued to be due to analogy with no-change irregulars (cut- cut) rather than due to the implicit generalization of a product-oriented schema. Connectionist models more generally demonstrate neighborhood effects whereby items that fall into densely populated neighborhoods, for example drink-drank and sink-sank, tend to be inflected similarly by virtue of analogy (which can mean they are especially resistant to overregularization errors).

One final type of model to consider is referred to as the parallel dual-route race model (Baayen, Dijkstra, & Schreuder, 1997). In this model, both whole inflected forms and decomposed stems and inflections are activated and race for selection. The higher the frequency of the inflected form, the more likely the whole form is to be used, not the decomposed stems. This model is thus similar to Bybee’s (2001) model we described previously but has a more specified processing architecture that includes steps for the activation of segments, the licensing of their composition, and the composition of their meaning. This architecture is proposed to explain adult comprehension. Consequently, it is difficult to make any strong predictions about errors of omission in child production. However, the pervasive storage assumed by the model would presumably predict fewer errors of omission for words whose inflected forms are of higher frequency. The logic here is that inflected forms that are of high frequency are more likely to be retrieved as a whole than to be formed from a stem and a separate inflection. Errors of omission would not occur if the whole form is retrieved directly but might occur if the stem is retrieved but then not combined with an inflection. The model also predicts that the inflected form of a word should be easier to process if that word is encountered relatively more often in its inflected form than in its uninflected form (when total stem + inflected form frequency is held constant). Taking the example of English nouns, the word cloud is often encountered in the plural form and is thus termed plural dominant, whereas ceiling rarely occurs in the plural and is thus is singular dominant. It was found that adults were quicker to process plural dominant plurals such as clouds than singular dominant plurals such as ceilings. Applying this reasoning to acquisition, we might predict that for words of overall equivalent frequency, those that occur more often in the inflected form will be less susceptible to errors of omission than those that occur more often in the singular (see also Bybee, 2001; Hay, 2001).

To summarize, the following six major factors have been identified as potentially affecting the rate of errors of omission in child English (Matthews and Theakston 2006, p. 1030):

1. The child’s ability to inflect productively (possession of a rule/schema).
2. The phonological similarity of the stem ending to a relevant inflection schema.
3. The phonological neighborhood of the item to be inflected (i.e., the number of other phonologically similar stem-inflected form pairs, e.g., sing-sang; ring-rang).
4. The regularity of the word (potential for competition effects).
5. The overall token frequency of the word.
6. The relative frequency of the inflected form to the stem.