is characterised by rules. Rules of grammar
, for instance, are many and permeate all aspects of language use. Without these rules, language would make no sense - after all, every language user needs to use a common linguistic system, else communication will fail, this being the goal
of language in the first place. It is not surprising, then, that based on the many observable rules and trends seen in language, models of various aspects of language have been constructed based around the language system of the brain
exploiting explicit rules.
One of the main focuses of such linguistic models has been in the rules governing English past tense morphology. The main issue here is that it is not a simple case of adding a particular morpheme (i.e., "-ed") to the end of every verb in the language to produce its past tense, since there are a significant number or verbs in English that do not conform to this regular pattern. Any model of English past tense formation must account for the developmental phenomenon over "overregularisation". This phenomenon manifests itself by a child producing correct past tense forms of verbs, then suddenly making errors by adding regular endings to irregular verbs. For example, children might suddenly use "goed" or "gived" after previously using "went" or "gave". There is then a phase of development where children will use incorrect forms (about 5% of the time) while also consistently producing correct regular forms, before correct adult usage is established. Another effect that models must account for is that children will automatically regularise novel verbs when using them in the past tense.
Converting verbs to their past tense forms is a process that obeys certain rules. The rules for regular verbs are simple, and even irregular forms have some rule-like quality to them, although rote learning is also an important strategy in dealing with them. Children's ability to automatically regularise novel verbs in the past tense has been cited as a key piece of evidence to suggest that children have acquired a rule to form the past tense which they apply and follow (e.g., Pinker, 1984). This process cannot be the simple recollection of a verb form acquired by rote, as the child has not heard the verb before (e.g., Berko (1958) found that children would regularise the nonsense verb "rick" to produce "ricked"). Children must surely apply a rule then.
On the basis of this and similar evidence, Pinker (1984) formulated a "dual mechanism" account of the formation of English past tenses. This model takes a verb stem as input and uses a rule-governed process to add the regular past tense "-ed" ending. However, there is another parallel mechanism which checks the input stem to see if it is an exception to the regular rule, in which case, it blocks the regularisation process and recalls the irregular form from some kind of storage location. This model explains the overregularisation effect in that during development after the rule process has been established, the initially weak memory engrams formed for the irregular exceptions are not always able to muster sufficient strength to recall them from memory to be used in the blocking procedure, thereby allowing irregular stems to be regularised. It also explains the fact that novel verbs are regularised due to the novel stem not being present in the memories of exceptions.
Pinker's model contains within it an explicit representation of a rule. The model is all well and good, but somehow its seems a little abstract to be an accurate representation of how the process occurs in the brain. Maybe it reflects the cognitive processes involved, but the question remains as to whether "rules" are represented in this explicit fashion.
There is another possibility provided by the connectionist approach, through interactive activation models of linguistic processes, such as the model of English past tense formation proposed by Rumelhart and McClelland (1986). In this model, lawful processes can be modelled by a mechanism without any explicit representation of a rule. Rumelhart and McClelland's model is able to learn paired inputs of a verb stem and its past tense form and associate the two by using a back propagation algorithm to adjust the weight matrix connecting three levels of nodes (input nodes, "hidden nodes", and output nodes) to improve the accuracy of the output to the input. Ultimately, the model is able to take just a verb stem as input and accurately produce its past tense form as output.
Pinker and Prince (1988) have written extensively in criticism of Rumelhart and McClelland, but although their criticisms highlighted flaws in Rumelhart and McClelland's model, they did not succeed in their goal of spelling the death of connectionism. The problems with the Rumelhart and McClelland model (such as those surrounding its use of "Wickelfeatures" in phonological representation, and the fact that the U-shaped development of correct past tense usage may have arisen because of the drastic increase in training input) have been overcome by subsequent models such as those by MacWhinney and Leinbach (1991) and Plunkett and Marchman (1993).
The connectionist models seem two have two main, and significant, advantages over rule-based models. Firstly, connectionist models have predicted a great deal of the observed characteristics of linguistic processes, and, not only that, have accurately predicted effects not yet observed (e.g., Plunkett et al.'s (1992) conectionist model of vocabulary acquisition, which predicted that there would arise a sudden and rapid increase in the number of words a child can comprehend, a prediction later confirmed experimentally). Secondly, rule-based models have yet to produce a workable model in the same way as connectionist models. In the words of MacWhinney and Leinbach (1991), connectionist models have "implementations", whereas symbolic models have only "principles". These two facts suggest that the connectionist approach is more likely to accurately model how processes occur in the brain. It also seems more biologically and neuronally plausible (e.g., Hebbian processes could mediate the changes in a biological weight matrix). These factors add credence to connectionism as a whole, and more specifically to its non-reliance on explicit rules.
But even if connectionism is eventually shown to be incorrect, it cannot be argued that it has shown that behaviours that are characterisable by rules, need not have explicit rule-governed processes underlying them. Language seems to be more about extracting regularities and patterns, rather than learning rules.
MacWhinney, B. and Leinbach, A.J. (1991) Implementations are not conceptualizations: Revising the verb learning model. Cognition, 40, 121-157
Plunkett, K., and Marchman, V. (1993) From rote learning to system building: acquiring verb morphology in children and connectionist nets. Cognition, 48, 1-49
Rumelhart, D.E., and McClelland, J.L. (1986) On learning the last tense of English verbs. In J.L. McClelland, D.E. Rumelhart, and P.R. Group (Eds.), Parallel distributed processing: Explorations in the Microstructure of Cognition, Vol. 2: Psychological and Biological Models (pp. 216-271), Cambridge, MA: MIT Press