DISCUSSION

This study provides some support for the predictions of the dual-mechanism model, but also some contradictions.⁸

First, the model predicts a quantitative difference between the regular and the irregular verbs. The regular verbs should all be suffixed, whereas irregular verbs should be increasingly suffixed as they move further away from the prototypical English irregular verb. As illustrated in Figure 1, all groups did indeed suffix more regular than irregular verbs, supporting one of the predictions of the model. This finding, confirming one of the results from Prasada and Pinker's (1993) experiment 3 with the modified experimental procedure used in this study, supports the first prediction of this experiment.

One goal of this research was to ascertain whether there would be group differences in production of the past tenses of these nonce verbs. In particular, the study was designed to ascertain whether there are any fundamental differences between English NSs (both adult and child) and NNS adults (prediction 3a). Figure 1 shows that there were group differences, but the locus of the difference was found between the NS children and the NS adults. Thus the fundamental difference in this study lies not between L1 and L2 but rather between the NS children and NS adults. Furthermore, there was no Group × Verb Type interaction, which indicates that all three groups of participants showed the same difference between regular versus irregular verbs. This evidence suggests that the issues raised by the dual-mechanism model are relevant to the L2 learning context, as these participants treated regular verbs reliably differently from irregular verbs. This finding also suggests that L1 and L2 learning are not fundamentally different in this respect, given that there were no differences between the NS and NNS adults. Thus, prediction 3a is not supported, as the NS children did not pattern with the NS adults. However, prediction 3b is partially confirmed, as there were no statistically significant differences between the NNS adults and the NS children. However, it cannot be fully supported because the NNS adults were not statistically different from the NS adults, either. Both the NNS adults and NS children had less exposure and experience with English than the NS adults. To accept this evidence as supportive of the dual-mechanism model, however, one would have to also accept that dissociations between regular and irregular inflectional features are a direct result of two separate systems. Such an assumption may be erroneous given the weight of evidence suggesting that dissociations can arise without having to invoke two separate processing-representational routes.

A further issue concerning potential group effects was whether they would be present on both types of verbs. Recall that, to generalize one's knowledge of an irregular past tense to a novel item, one must have already stored in associative memory a representation of some form of a past-tense verb. For example, if one is to generalize the past tense of swing to the nonce verb spling (and hence produce splung as a possible past tense), one must have swung stored as a past tense in associative memory. Therefore, participants who vary in terms of how much exposure, familiarity, or both they have had with the language in question might reasonably be expected to perform differently on the set that is influenced by associative memory (i.e., the irregular verbs). If, however, all participants have knowledge of the rule of forming the past tense by adding [-ed], then one might reasonably expect to find no group differences on the regular set of nonce verbs.

The crucial issue regarding the potential group effects and the dual-mechanism model concerns the interaction between regular and irregular items. Assuming all three groups of participants had knowledge of the [-ed] rule, then any group differences should have only been apparent on the irregular items. Figures 3 and 4 illustrate that on both the irregular (Figure 3) and regular (Figure 4) verbs, group effects were evident. Specifically, NS children produced more suffixed verbs on both the irregular and regular verbs than the NS adults. Recall that the NS children in this study were between 8 and 9 years old. One might reasonably conclude, therefore, that they had had sufficient exposure to English and sufficient knowledge of the English past-tense system to enable them to perform this task. Therefore, all participants in this study would have had sufficient knowledge of the [-ed] rule but also would have had knowledge of a variety of different irregular verbs.⁹

Given that the NS children distinguished themselves from the NS and NNS adults on suffixation, it would seem that age was an influencing variable as to how participants suffixed the verbs in this experiment.

One possible explanation as to why NS children suffixed more than the NS adults might come from what Pinker and Prince (1994) referred to as the blocking principle. In attempting to understand how the regular and irregular systems interact with one another, Pinker and Prince suggested that retrieval (from memory) of an irregular past form blocks the application of the rule. They also suggested that in real time, participants would search through memory for a past irregular stem but, at the same time, feed the stem into the regular rule mechanism. A match from memory then blocks the rule application. However, if there is no match from memory, the rule-application mechanism proceeds, and the verb is suffixed. This blocking principle might account for the NS children suffixing more than the NS adults. Pinker and Prince noted that “children's memory traces are not strong enough to guarantee perfect retrieval” (p. 344). That is, their memory traces and connections between stem and past forms in associative memory will not be as strong as adults. A consequence of weaker connections is that successful retrieval from memory will fail more often in children than in adults. According to the blocking principle, then, if there are more failures at retrieving an item from memory, there will be a commensurate increase in the number of suffixed verbs.

There are at least two problems with this suggestion. The first is that in discussing weaker memory traces in children, Pinker and Prince (1994) were referring to much younger children than the ones tested in this study. Indeed, Pinker and Prince were discussing the tendency to add [-ed] to either irregular stems (e.g., break → breaked) or to already inflected irregular past forms (e.g., eat → ate → ated) in children around 3 or 4 years old. Although it may be true that children as young as 3 or 4 have weaker associations between irregular stems and past forms, it is less obvious that children between 8 and 9 will also show such weak connections in memory as to cause greater suffixation.

A second problem with the blocking-principle account relates to the NNS performance. There was no statistically reliable difference between the NNS and NS adults (or the NNS adults and the NS children) in terms of number of verbs suffixed. One might have predicted NNS adults to suffix more, given that they had spent the least amount of time learning and using English and hence could arguably have had the weakest memory traces for these forms. If this were the case, one would have expected at least a difference between the NS adults and the NNS adults. One would have to suggest, counterintuitively, that something inherent in being a child (of 8 or 9 years) causes weak memory traces and thus results in greater suffixation, despite the fact that arguably the NNS adults actually had less exposure and familiarity with English than the NS children.

Another possible explanation for greater suffixation by the NS children than the NS adults might be related to attentional processes. The individual items may have been more salient for the NS adults, the NNS adults, or both than for the NS children. Many researchers have argued that L1 learning is largely implicit, which suggests children are not in any conscious way attending to the specific linguistic forms. Contrast that idea with Schmidt (1994), who argued that language learners must “notice” a particular linguistic feature to learn it. In the context of this study, it is likely that all participants would have noticed that the verbs presented to them were odd in some specific ways. Indeed, there may be an inherent explicitness to a task that requires participants to find an analogy on which to base a new past tense. If this is true, and if children of 8–9 years of age have weaker attentional processes, then they might be expected to suffix more.

The roles of attention and memory have been considered to be powerful determinants on learning in general (e.g., Cowan, 1997) and learning language in particular (Ellis, 1994, 2001; Gathercole & Baddeley, 1993). Early theories of attention suggested that if input is not attended to, then it will not be encoded so as to be recalled later on (e.g., Broadbent, 1958)—an idea similar to Schmidt's noticing hypothesis (1994). In the past-tense generalization task of this study, participants have to hold the nonce verb in their working memories as they search through their longer term associative memories to determine if there is a similar form on which to base an analogy. If children have weaker attentional processes, it may be harder for them to hold and process the nonce verb in working memory. If this were true, it may be more difficult for them to find an appropriate lexical item in their associative memories on which to base their analogies. Such a problem could then lead the child to suffix the verb as they would be less likely to find an analogous item in their associative memory. There is some evidence that older children are able to search through their memories faster than younger children (Cowan, 1992; Cowan et al., 1994), which suggests a developmental progression in memory retrieval processes. If the adults were able to search through their memories faster than the children in this experiment, then they would have been more successful at finding the appropriate analogous representations and would have suffixed less than the children.

It is difficult to determine at this stage whether such an explanation could account for increased suffixation by the children. Further research is needed to identify more precisely the role of attention, memory, and language learning in both L1 and L2 learning contexts for both children and adults.

An additional focus of this study was to examine the second prediction that similarity effects would only be manifest on irregular items. Recall that the dual-mechanism model predicts that similarity should only influence irregular verbs. Figure 3 illustrates that, as predicted, there was a similarity effect on the irregular verbs because more irregular verbs were suffixed as a function of distance from the prototypical irregular verb. Interestingly, Figure 4 also indicates a similarity effect on the regular verbs. The effect is in the opposite direction in that fewer regular verbs are suffixed as a function of distance from the prototypical.¹⁰

This similarity effect for the regular verbs, which holds for all participants, does not support the dual-mechanism model. The methodology used in the present experiment was not identical to that of Prasada and Pinker (1993), as the nonsense verbs in this study were not introduced within a semantic context as they were in Prasada and Pinker's study. Ramscar (2002) has identified how important semantic context can be in biasing a participant to inflect a nonsense verb using either a regular or irregular form. The focus of this study was to identify the extent to which the phonological characteristics of the nonsense verbs would influence the participants to derive either a regular or irregular inflected verb form. It is possible, therefore, that the unexpected similarity effect found here on the regular nonsense verbs is due to the fact that the participants in this study had to rely entirely on phonological characteristics of the nonce verbs, whereas in the original study, Prasada and Pinker's participants also had some semantic context.

If the differences between the original Prasada and Pinker (1993) study and this one are due to this methodological difference, it would strongly suggest that phonological properties (and importantly, participants' perceptions of the similarity between real English and nonsense verbs) lead participants to inflect verbs in ways that are inconsistent with the predictions of the dual-mechanism model. The findings in the present study are important because they provide an opportunity to identify how the phonological similarity alone (without recourse to any possible influence from semantics) influences generalization of English past-tense inflection.

Similarity is assumed not to influence processing of regular verbs, as suffixation is in some sense a blind computational procedure. At the very least, as Prasada and Pinker (1993) suggested, processing of the regular verbs does not depend on similarity in the way it does for irregular verbs. Prasada and Pinker claimed that generalizations of regular inflections are made on the “basis of a process that is not dependent for its success on how similar the phonological properties of the stem are to previously encountered verbs undergoing that morphological alternation” (p. 28). This suggests that, although similarity might influence processing of regulars, it does not have to, whereas it does have to for the irregular items. Interestingly, Prasada and Pinker reported the mean suffixation for the regular prototypical, intermediate, and distant items as 9.5, 9.2, and 8.9, respectively. Although this difference in means was not significant in their study, the pattern is in the same direction as the significant effect found in this research. The results of the present experiment would then indicate that, without semantic context, similarity does influence both regular and irregular verb formation. To reconcile these results with the dual-mechanism model, one would have to suggest that under some circumstances (no semantic context), similarity can influence regular as well as irregular verbs, but in others (semantic context), similarity does not influence past-tense generalization on regular verbs. However, such a suggestion directly contradicts many claims Pinker (1999) has made concerning associative factors influencing only irregular items and is neither predicted nor explained in current accounts of the dual-mechanism model.

Another approach that might help explain these findings is found in Ullman (2001). Ullman's declarative-procedural model claims that the lexical and grammatical systems are distinct systems (as the dual-mechanism model). Ullman's model also maintains that one component (declarative) adheres to an associative memory structure, whereas the other (procedural) underlies a symbol-manipulation system (again similar to the dual-mechanism model). However, the declarative-procedural model differs from the dual-mechanism model in that each of these two systems is subserved by its own set of more domain-general, nonlinguistic functions or mechanisms. Ullman examined neurological evidence from patients with aphasia and other forms of language impairment to support these claims.

The research presented in this paper was not designed to test Ullman's (2001) declarative-procedural model. However, it is interesting to examine some of the predictions of Ullman's model within the context of this paper, especially because he makes some specific predictions about L2 learning. Ullman stated that markers of associative memory (like frequency) should be found for memorized items (items ostensibly in the declarative system) and not for those items computed by grammatical rules. Similarity (as discussed previously in the introduction) is also a marker of associative memory, and therefore within the declarative-procedural framework, we should expect that similarity should primarily influence the production of the irregular nonsense past-tense forms. The data presented in this paper illustrate that similarity does influence the irregular past-tense productions but also influences the regular productions. Certainly, the pattern of the similarity effect for both the regular and irregular nonsense verbs was different, and whereas the presence of a similarity effect on the regular forms is strictly prohibited by the dual-mechanism model, perhaps such an effect could be somehow accounted for in the declarative-procedural framework through recourse to the nonlinguistic factors that subserve the computational system. More research clearly needs to be done on this issue to test and, more specifically, identify how these nonlinguistic factors relevant for both the declarative and procedural systems affect specific tasks such as the one described in this paper.

Additionally, Ullman (2001) predicted that L2 learners who have learned their L2s later in life are more likely to rely on the declarative (associative) component for all aspects of linguistic processing. In other words, the procedural component seems most affected by being an older language learner. This study does not test for this assumption, nor was it intended to. However, this claim could explain why the NNS adults manifest a similarity effect on both the regular and irregular verb forms. If the NNS adults rely more heavily on a system that is more influenced by factors such as similarity and frequency, one might expect them to show such an effect on items typically under the jurisdiction of the computational or procedural component. Although this idea is helpful in understanding the similarity effect for the NNS adults, it does not necessarily or directly explain the similarity effect on the regular forms found in the NS adults and NS children (unless the nonlinguistic factors underlying the computational-procedural system are responsible). In any case, Ullman's work presents a new and interesting framework within which to test these ideas. For the present, however, as this study was not designed to test the predictions of Ullman's model, we have to be cautious in concluding the extent to which these results support or refute some of the claims of the declarative-procedural model.

Prasada and Pinker (1993) argued that a connectionist theory would have trouble accounting for the regular system but can account for processing and representation of the irregular system. MacWhinney (1993) has shown, however, that a connectionist model can account for both the regular and irregular systems and accurately simulate both the regular and irregular nonce verb generalizations. Therefore, models that operate on purely associative processes can show the same kinds of dissociations presented and discussed in Prasada and Pinker.

We have already seen that behavioral dissociations between regular and irregular items of the kind cited by Prasada and Pinker (1993) and Pinker and Prince (1994) as supportive of the dual-mechanism model can be accounted for by single-route associative accounts. We have also seen in this study that some of the key predictions of the Prasada and Pinker (1993) study were not replicated. We can provisionally conclude, then, that an associative account of processing and representation is the better candidate at explaining these results for both L1 and L2 speakers. Importantly for the L2 learning context, the results of this study indicate that NNS adults, like NS children and NS adults, use similarity to make their past-tense generalizations for both irregular and regular verbs. The lack of reliable differences between the NS adults and the NNS adults suggests that these two groups use similar processes both in terms of the number and type of verbs they suffixed and in terms of the type of irregular nonsuffixed responses they produced. These results also parallel those presented in Lee (1994).

As previously discussed, Lee (1994) pointed out a number of ways in which the stimuli developed in Prasada and Pinker (1993) were problematic. Lee developed what he considered more “principled” pseudo-irregular and regular verbs and used them in a past-tense production task—in the form of a cloze passage—to both NSs and NNSs. His data matched closely with those discussed in the present study. First, Lee found that similarity influenced the irregular forms in that more items were suffixed as a function of distance from the prototypical. For Lee's nonce regular verbs, there was a marked decrease in the number of suffixed distant regular verbs relative to the prototypical and intermediate regular ones. Thus, Lee also found a similarity effect on the regular items—with stimuli that were putatively more realistic and where the distance between the three categories of verb was varied consistently. As in the case of the results reported here, the adult NSs and NNSs in Lee's study produced similar patterns of responses, which suggests that both groups of participants use analogy as a basis for generalizing their knowledge of past tense.

The issue of the reliability of the items used in this and Prasada and Pinker's (1993) studies remains problematic, however. One could argue that the reliable effects found in this study were due to characteristics of specific items. In examining the items closely, we see for example that plip is considered a prototypical regular verb, whereas blip is considered a distant irregular one. If the distant irregular verbs were like the prototypical regular ones, then we might expect subjects to suffix the distant irregular forms more, which is what they did in the present and Prasada and Pinker's studies. This problem with the items could have potentially biased the results in the direction of supporting the dual-mechanism model (again, assuming that a dissociation between regular and irregular items is the kind of evidence needed to support the model). The purpose of the present study was to extend the original Prasada and Pinker study to a different sample of participants; therefore, it was important to use the same stimuli. It is interesting to note that Lee (1994) also found the same pattern of results even after attempting to resolve some of the problems associated with constructing the items. Item effects, however, do not pose a problem for an explanation based on principles of associative learning but present difficulties for a symbolic rule-based account. If the findings of this and past work are due solely to features of individual items, then it precludes the possibility of a blind computational, symbolic process at work.

The results of this study alone do not portray a sufficiently clear picture to warrant unequivocal support of a single-route associative system, either in L1 or L2 processing. However, when these results are taken in the context of other research investigating similar issues, it becomes clear that the dual-mechanism model as described by Pinker and Prince (1994) cannot be supported. For example, as discussed earlier, empirical work with humans investigating the dual- versus single-route hypotheses has indicated that dissociations between regular and irregular inflectional features could be accounted for within a single associative account. Ellis and Schmidt (1998) found that adult participants learning morphosyntax show frequency effects on the regular items early on. It is only later on the learning curve that the effect of frequency on the regular forms diminishes relative to the irregular forms. Ellis and Schmidt suggested that a ubiquitous law of learning—“The Power Law of Practice” (see Anderson, 1982)—can account for this putative change in behavior. That is, the extent of improvement on a particular task decreases as a function of an increase in frequency or practice on the task. There are considerably more regular than irregular verbs in English; thus, the regular morpheme is by far the most frequent past-tense form in English. Any difference, therefore, between regular and irregular features could be a direct result of the increased frequency and amount of exposure that learners have with the regular features.

Other research with human participants has indicated that they can sometimes exhibit putative dissociations between different grammatical categories. For example, Bates, Friederici, and Wulfeck (1987) and Blackwell and Bates (1995) have discussed how otherwise neurologically sound participants manifest dissociations on agreement marking relative to other linguistic features under conditions that put severe burdens on the working memory systems. The important point here is that dissociations may not always indicate fundamental differences at the learning or representational levels, or both.

Additionally, the work reported in the present research suggests that processing of irregular items may not even be dissociated from processing regular forms (i.e., even if a dissociation means separate processing and representational systems). Similarity influenced how participants generalize their knowledge of past tenses to novel items. The direction of the differences varied in that similarity caused fewer items to be suffixed in the regular set and more items to be suffixed in the irregular set as a function of distance from the prototypical. The point remains, however, that similarity influences both category of verbs. Beck (1997) also reported that frequency influences processing of both the regular and irregular sets of verbs.

Given this evidence, one can reasonably question the claim that there are two dissociable learning systems underlying language learning and representation for both L1 and L2 learning. That is, there are a number of indications that a more parsimonious account invoking a single set of associative learning mechanisms could account for the development of inflectional properties of language.