2.2 The Dual-Mechanism Approach

Unlike the single-mechanism approach which takes extreme positions that regular and irregular forms are processed and represented by either only the rule system or only the memory system, the dual-mechanism approach takes a hybrid position that regular and irregular forms are processed and represented separately. It further posits that regular forms are composedin real time by rule which concatenate suffixes to stems into morphological constituents and decomposed by rule into stems and suffixes, whereas irregular forms are full-form representations stored in and retrieved from the mental lexicon through an associative memory. Additionally, successful retrieval of the irregulars blocks the application of rules. For the dual-mechanism approach, agenerative suffixation rule provides a necessary and sufficient account of generating the regular and novel forms; but since this procedure cannot deal with exceptions to the rule, an additional associative memory mechanism is proposed.

The dual-mechanism approach began with Pinker and Prince's (1988) critic of Rumelhart and McClelland's (1986) connectionist model of past tense learning. Since then, a number of variants of the dual-mechanism models have emerged, such as the Dual-route Model (Caramazza, Laudanna & Romani, 1988), the Morphological Race Model (Baayen, Dijkstra & Schreuder, 1997), the Words and Rules Model by Steven Pinker and his collaborators as well as the Declarative/Procedural Model by Michael Ullman and his collaborators. The last two models have won greater popularity and in the following sections these two models will be introduced in details.

2.2.1 The Words and Rules Model

The Words and Rules (WR) Modelis a psycholinguistic modeling of the rule-and-memory linguistic theory. It is descended from the lexicalist theories (e.g., Jackendoff, 1975; Lieber, 1980), which recognize that many morphological phenomena are neither arbitrary nor fully systematic and productive. WR claims that two complementary systems coexist in how our brain processes and learns a language. One is the combinatorial (rule-based) system that generalizes over symbolic categories, such as N (oun), V (erb), and treats all members of a given category equally. The other is the associative system that extracts probabilistic contingencies from the input data, such as frequency distributions, similarity clusters. It shares some properties of the associative memory in the connectionist models and has the ability to generalize by analogy. The key property of the combinatorial system is that it suppresses differences between individual examples and treats all members of a group or class equally. By contrast, the associative system generalizes on the basis of resemblance to stored examples. The key predictions of WR are that irregulars should have the psychological, linguistic and neuropsychological signatures of the lexical memory, while the regulars will often have the signatures of grammatical processing, and that regular inflections should be appliedas the default treatment whenever memory fails to supply the target form for a particular irregular.

The Words-and-Rule Model is motivated by the observation that the regular/irregular distinction, inparticular the distinction between mental lexicon and mental grammar, is an epiphenomenon of the design of the human language faculty. Irregular forms are just words, acquired and stored like other words, but with a grammatical feature like[+Past tense]incorporated into their lexical entries. Regular forms, by contrast, can be productively generated by rules, just like phrases and sentences. Moreover, when one produces an irregular form, he not only has to resort to memory but also must inhibit the rule like “adding-ed”, so he doesn't say bringed or broughted. That is, a stored inflected form of a word blocks the application of the rule to that word (e.g., brought represses bringed). Elsewhere the rule applies by default: it concatenates suffixes like-ed with the symbol representing a category like “V”, and thus can inflect any word categorized as a verb. This is the so-called Blocking Principle. One possibility for blocking is that when one needs to utter a past tense form he first scans his list of irregular verbs to see if it is there, and if it isn't, he turns on the rule. This predicts that the slowest retrieval of irregular verbs should be faster than the fastest retrieval of regular verbs. Another possibility is that words and rules are accessed in parallel, that is, at the same time (see Figure 2-2). As one plans to utter a verb in the past tense, he simultaneously looks up the word in memory and activates the rule. An inhibitory link runs from the memory box to the rule box, which gradually slows down the rule as evidence for a match is found, and eventually turns it off.

As to how the blocking principle is learned by children, Pinker (1999: 197) explains that there is no plausible scenario for how children might learn such an abstract principle in the course of acquiring their first language. Instead, he proposes, blocking might be built in to the circuitry that drives language acquisition—what Chomsky calls Universal Grammarand what he calls the language instinct.

Figure 2-2 The Words and Rules Model

(Adapted from Pinker & Ullman, 2002: 457)

The working principle of the WR model is shown in Figure 2-2. WR posits a parallel-race model. When a word (e.g.,talk) must be inflected, the lexicon and grammar are accessed in parallel. The stem and the suffix-ed are retrieved from the lexicon and then access into grammar. There a rule will apply to it, which concatenates the past tense suffix-ed to the stem. Thus, the suffixed form (talked) of the stem is generated. If an inflected form for a verb exists in memory, as with irregulars (e.g.,told), it will be retrieved directly from the lexicon. A signal indicating a successful match blocks the operation of the grammatical suffixation process via aninhibitory link from lexicon to grammar, inhibiting the generation of telled. But sometimes memory fails to supply the irregular forms or no inflected forms in the memory are matched, then the suffixation rule will apply as the default treatment and the grammatical processor concatenates the appropriate suffixes with the stems, generating regular forms. Hence, the phenomenon of overregularization will arise. A stored irregular may be inaccessible for several reasons, such as very low frequency, a lack of a similar form that could activate an analogy, exocentric structure, novelty of the form in childhood and damage to the neurological substrate of the lexical memory.

In summary, the Words and Rules Model holds that irregular forms are stored inand retrieved from the lexical memory with the property of pattern associators, whereas regular forms are generated in realtime by a distinct symbol-manipulation system which applies a suffixation rule to the stem. Retrieval of an irregular blocks the application of rules, but when an irregular is not successfully retrieved, rules will be applied by default, resulting in overregularization.

But we should avoid some misunderstandings about this model. Firstly, according to Pinker and Ullman (2002: 458), it does not posit that regular forms are never stored, but it only posits that they do not have to be. Whether a regular form is stored or accessed for retrieval depends on word-, task-and speaker-specific factors. For instance, they argue that regular forms of doublets, such as dived/dove, dreamed/dreamt, must be stored to escape blocking by the irregular forms. Thus, regular doublet forms should show a strong frequency effect. The same is true for regular forms of verbs that resemble irregulars like blinked, winked, showed, played and glided, because these forms must overcome a partial blocking effect exerted by the similar irregulars. Secondly, the WR model is not just a connectionist pattern associator glued onto a rule system. The lexicon has associative properties similar to pattern associators, but lexical entries have structured semantic, morphological, phonological and syntactic representations which are not implemented in pattern associators. Finally, Pinker (2006: 232) emphasized that the word in WR refers to language chunks of any size-morphemes, phrasal idioms or even proverbs-that are not compositional in nature but have to be memorized, while the rule in WR refers more broadly to any productive and computational operations on symbolic structures.

2.2.2 The Declarative／Procedural Model

Given that words are like facts (semantic knowledge) and events (episodic knowledge) in being arbitrary and the lexical memory is a subdivision of the declarative memory storing facts, events and arbitrary relations, and that grammatical rule are like skills in requiring coordination of procedures in real time and grammatical processing relies on the procedural system, which underlies the learning and processing of motor, perceptual and cognitive skills involving sequences, Ullman and his colleagues have recently extended the Words and Rules Model to a hypothesis about the neurocognitive substrate of lexicon and grammar, named the Declarative/Procedural (DP) Model (Ullman, 2001a, 2001b, 2004; Ullman et al, 1997; see Figure 2-3). It posits that the lexicon/grammar distinction in languages is tied to the distinction between two brain memory systems and proposes that the processing of one's native language involves these two different brain systems. One is the declarative memory which is responsible for a lexical store and retrieval of memorized words (at lease simple words with noncompositional arbitrary form-meaning pairings) and is rooted in a network of specific brain structures including medial temporal and prefrontal cortical regions, the hippocampus in particular. The other is the procedural memory with combinatorial rules and rooted in a network subserved by basal-ganglia circuits connected largely with frontal cortex. Learning and remembering new, and controlling well-established, motor and cognitive skills in the procedural memory systemare largely implicit. Different from the WR model and the connectionist models, only the DP model predicts associations in processing and representation of irregular forms, non-compositional lexical items, facts, events with temporal-lobe structures as well as associations between regular forms, aspects of syntax and other domains of grammar, motor and cognitive skills, and frontal/basal-ganglia structures.

According to the DP model, morphological transformations can be computed by either of two components with distinct cognitive, computational, and neural bases. One is the declarative memory of distributed representations by which the transformations' phonological and conceptual-semantic mappings are learned, stored, and computed. It is partially productive and may generalize some patterns to new forms, but its extent of productivity remains unclear. Like the connectionist models, rules in this memory system are only descriptions of patterns in the language. The other component is a procedural memory system which subserves those morphological transformations that are fully predictable and productive. Its rules specify any sequential and hierarchical structure of regular forms and compute morphophonological transformations in real time by symbol manipulation. It concatenates the phonological forms of stems and suffixes (e.g., walk + -ed → walked; rat +-s → rats).

Figure 2-3 The Declarative/Procedural Model

(Adapted from Pinker & Ullman, 2002: 457)

The working principle of the DP model is as follows. Like the WR model, the computation of amorphologically complex form in the DP model also involves the parallel activation of the two systems. The declarative memory system stores and retrieves a form, while the procedural system computes a rule product in real time. As the retrieval proceeds in the declarative memory, a signal indicating the success or failure of the retrieval of a particular form is sent to the rule-processing system. This successful signal prevents the procedural system from carrying out computation. For instance, the successful retrieval of “brought” blocks the computation of “bringed”. If a memorized form is not retrieved, then the rule will apply as the default and lead to overregularization. On the other hand, the successful computation of a form by the procedural system should inhibit the memorization of the form in the declarative memory, thus reducing the likelihood of memorizing regular forms. Any regular form, however, can be memorized in principle. The likelihood of its memorization should increase with factors like frequency. That is, the high-frequency regulars may also be memorized and stored in the declarative memory.

Finally, according to Ullman (2004: 247), although the declarative memory system and the procedural memory system function differently, the DP model assumes that there may be interactions between these two systems. Firstly, the procedural memory builds complex structures and learns rule-governed patterns by selecting lexical items from the declarative memory. Secondly, the same or similar types of knowledge may be acquired by both systems, at least in some cases. And the rapid declarative storage of sequences of lexical forms may provide a database from which grammatical rule can gradually and implicitly be abstracted by the procedural memory system. Thirdly, access to a stored representation with similar mappings to the one which could be combined compositionally by the procedural memory (e.g.,an irregular vs a regular past-tense form of the same stem) would block completion of the latter computation. Moreover, learning in one system may depress functionality of the other. It is possible that enhancing the estrogen function in medial temporal lobe structures may result not only in an improved declarative memory function, but also in a suppressed procedural function.

2.2.3 Summary

The dual-mechanism approach assumes both a computational component containing rule and an associative memory system with some properties of the connectionist models. Within this approach, regular forms are computed by rule of adding suffixes to a variable standing for the class of the stems, whereas irregular forms are memorized pairs of words and the linkage between the mappings is stored in an associative memory in the way of analogy. Furthermore, regular inflectionsare rule-driven so that speakers can apply it whenever memory fails. A stored inflected form, however, blocks the application of the rule to that stem. In essence, the dual-mechanism models disagree with the connectionist models as to whether mental computation requires the use of symbols and diverges with Generative Phonology as to whether a memory system is required to store and retrieve the irregulars.