3.1 Child First Language Acquisition

Earliest research into the processing and representation of regular and irregular forms concentrated on the evidence from child first language acquisition, mainly by looking at how children learn the English regular and irregular verbal system. The overwhelming studies in this field focused on the conspicuous phenomenon of overregularization like such as breaked, holded, comed, goed, etc.

Many studies (e.g., Ervin & Miller, 1963; Cazden, 1968) show that for months before the appearance of overregularization in child English, all the overtly inflected irregular past tense forms are correct but regular verbs are left unmarked most of the time. If both the irregulars and regulars are memorized, as the connectionist models advocate, regular verbs should be correctly inflected as well at this moment since the irregulars are correctly inflected. If both the irregulars and regulars are generated by rules, as the Generative Phonology advocates, both the regulars and irregulars should be left unmarked if children have not acquired the past tense suffixation rule or they both should be correctly inflected if children have acquired the past tense suffixation rule. Anyway, dissociation of the regulars and irregulars in child English acquisition seems to contradict the single-mechanism approach, but conform to the dual-mechanism approach.

When children begin to use some inflections on stem verbs, the overregularization arises. Later when the transitional phase comes to an end and children use inflections skillfully enough, overregularization disappears. This is the so-called U-shaped development in child language acquisition. It seems that children learning English move from an early stage of rote learning of individual past tense forms to the induction of rule-based representations. The overregularized verbs cannot be accounted for by some form of imitation or Skinnerian reinforcement procedures, because children would never be exposed to these forms in the input from adults. Their occurrence must be the consequence of children's induction of linguistic rules and their overapplication of rule to irregular verbs. Hence, the connectionist models cannot account for overregularization because it involves the application of rule while the connectionist models deny the existence of rules. Berko (1958) found that even preschool children could apply the regular suffix-ed to novel verbs like rick. And Pinker (1999: 196) also showed the existence of rule with overregularization by recording a dialogue between a child and an adult. After the child's speech “My teacher holded the baby rabbits and we patted them” was corrected by the adult “Did you say your teacher held the baby rabbits?”, the child answered “Yes”. But when then questioned “Did you say she held them tightly？”, the child still answered “No, she holded them loosely.”

The connectionist models (e.g., Rumelhart & McClelland, 1986) argue that the overregularization phenomenon arises because children become temporarily overwhelmed by the regular pattern due to an influx of regular verbs in the child's input. ButMarcus (1995) showed that the onset and rate of overregularization errors in English-speaking children did not correlate with changes of the number or proportion of regular verbs used by parents. Moreover, the input of verbs should be counted by tokens, not types. Although children have the input of many more regular verbs, irregular verbs are all high in token frequency and remain predominant in children' verb input. Based on Pinker (2006: 228), the rate of increase in regular verbs is, in fact, negatively correlated with overregularization in children's speech over time, and children's vocabulary spurt, which Rumelhart and McClelland attributed to the onset of overregularization, actually occurs a year before the first overregularization errors. Furthermore, the connectionist's account cannot explain why children later continue to have a large input of regular verbs, but they do not overregularize irregular verbs and overregularization comes to an end. It should be that when overregularization appears children are undergoing the abstraction of the past tense suffixation rule. Once they complete the abstraction of the past tense suffixation rule, they will know that this rule can be applied to only regular verbs, and overregularization will disappear. In a word, overregularization implies that rule are indispensable for language acquisition. Marcus et al. (1992) added that children have not lived as long as adults and their exposure to the irregulars is not as much as that of adults. Consequently, their memory traces of the irregulars are not as strong as those of adults, thus their retrieval is less reliable. That is why overregularization happens to children, not adults. Therefore, as for children's acquisition of the irregulars, overregularization is the result of insufficiently reinforced memory traces and the application of rule as the default treatment, and it disappears when children hear them often and remember them firmly enough. Thus, the overregularization in child language acquisition lends strong support to the dual-mechanism approach.

Some studies, on the other hand, have tried to argue for the connectionist models by showing that the use of regular forms is also influenced by frequency in child first language acquisition. For example, Stemberger and MacWhinney (1986) found that the rate of children's failure to add the appropriate inflections is higher for the regular forms with lower frequency than those with higher frequency. Moreover, Stemberger and MacWhinney (1988) investigated errors on the intended regular and irregular pasttense forms in spontaneous speech and a production task. They found frequency effects on regulars in the production task, but not in spontaneous speech. However, a big problem with their studies is that the stem frequency was not controlled. Since stem frequency and the frequency of inflected forms are highly correlated, their results could also be explained by the differences in stem frequency.

Marchman (1997) also reported frequency effects for both regular and irregular verbs in a past tense elicitation task given to children. And again, the stem frequency was not controlled. Moreover, the children in Marchman's study were quite young, with the youngest of 3 years 8 months and a mean of 7.5 years. Van der Lely and Ullman (2001) found that 6-year-olds, but not 7-or 8-year-olds, showed frequency effects for the regulars. It is very likely that the children studied by Marchman showed frequency effects for the regulars simply because they were young and still relied to a substantial extent on the memory system. In addition, the stimuli in the studies of both Stemberger and MacWhinney (1988) and the Marchman (1997) were almost exclusively “inconsistent” the regulars whose stems were phonologically similar to the stems of one or more irregulars (e.g., glide-glided vs ride-rode, slide-slid). Therefore, they and their similar-sounding neighboring stems do not follow a consistent pattern. According to the dual-mechanism theories, the inconsistent regular past tense forms are likely to be memorized; otherwise generalizations computed in the associative memory might lead to nonexistent forms like glid or glode, which could block the computation of the regular form glided. Ullman (1993, 1999) did find that “consistent regulars”did not yield any frequency effects, while “inconsistent regulars”did show frequency effects, even partialing out stem ratings.

Some other studies imply that irregular patterns are stored in an associative memory, but that regulars are not. For instance, Marcus et al. (1995) found that the more often a parent used an irregular form, the less often the child overregularized it. Besides, verbs were protected from overregularization by similar-sounding irregulars, but not attracted to overregularization by similar-sounding regulars. Similarly, Bybee and Slobin (1982) found that children made errors more often for the irregulars that their parents use less frequently and they made fewer overregularization errors for the irregular verbs which fall into families with more and higher frequency members.

There are also some other studies concerning the children's processing or acquisition of regular and irregular forms. For instance, in an experiment Gordon (1985) found that 3-to 5-year-old children responded with mice-eater 90% of the time when they were asked who knew the word for a “monster who eats mice”, but with rats-eater only 2% of the time when they were asked who knew the word for a “monster who eats rats”. Adults do not use such compounds often, so children cannot learn this constraint from adults. One possible explanation is that this constraint is a consequence of the inherent organization of the children's grammatical systems. Plunkett andMarchman (1991) simulated past tense development in children learning English as L1, using a model in which the input approximated the type frequency of irregular verbs in English. Their connectionist model produced the same kinds of past tense production errors as those of English-speaking children. Furthermore, the model's learning curve followed the same U-shaped developmental function characteristic of English-speaking children's past tense development. Plunkett and Marchman argued that the variable frequencies of the different verb types in the training setting led the model to simulate children's patterns of development. In a pasttense elicitation task given to children, Marchman (1997) reported that errors on the irregulars were, surprisingly, not significantly related to the number of the irregulars with phonological similarity. By contrast, errors on the regulars were affected by the number of regulars with phonological similarity. That is to say, the similarity effects were found on the regulars, but not on the irregulars, which is not consistent with either the single-mechanism models or the dual-mechanism models. However, as Ullman (1999: 52) pointed out, “this surprising contrast must be treated with caution, because no statistics were reported for the difference for the regulars, and stem access was not controlled for, so any similarity effects of stem forms were not factored out”. Finally, Jensvoll (2002) studied English/Norwegian bilingual children's acquisition of past tense and found that overgeneralization errors of the English and Norwegian default verb class do not depend on phonological similarity, while those of the other verb classes can be interpreted as being based on phonological similarity. Such results are consistent with the dual-mechanism approach.