Modularity and connectionism

Cognitive psychology has a long history of invoking the concept of modularity to explain various mental functions. In sum, this concept assumes that the separate components of a modular system are autonomous, i.e. they will remain “functionally intact when other components of the same system become corrupted” (Parkin 1996:4). There are several modular accounts of language function. For example, Lesser and Perkins (1999) propose three stages in listening for meaning: (1) phonological analysis of the string of sounds heard, (2) recognition of the word as a familiar one which is, therefore, available in the ‘phonological input lexicon’, and (3) semantic access, and the word processed for meaning (Figure 1). (See also Ellis and Young (1988) and Levelt et al. (1991) for similar modular accounts of speech production.)

Figure 1. Listening, looking and reading for meaning (Lesser and Perkins, 1999:14)  

Much neuropsychological evidence for separable modules of language function comes from an investigation of disorders observed in aphasic patients. Again, the number of disorders described is numerous. However, by way of illustration, three are summarized in Table 1.

The identification of so-called discrete disorders such as those listed in Table 1 are used to infer separable components, each possessing different processing resources and carrying out their own operations in isolation from what is taking place elsewhere in the system.

It is, therefore, possible to examine the clinical features of Case Study A (see Appendix 1) in terms of such a modular scheme. The patient’s auditory comprehension is reported as good/reasonable. This indicates that phonological analysis, the phonological input lexicon and access from this to the semantic system are probably intact. It also infers that the patient does not have a central semantic deficit. Fluent speech and poor repetition suggests impairment of the phonological assembly buffer. There are three routes for repetition of words: (1) a direct link from the phonological analysis system to the phonological assembly buffer, known as phonological input-to-output conversion (see Parkin, 1996:32), (2) from lexicon to lexicon, and (3) from phonological input lexicon via semantics to the phonological output lexicon. As the direct route bypasses the lexicons, it allows repetition of nonwords. Case Study A, however, provides insufficient information regarding the nature of the repetition difficulties. If repetition of nonwords were more difficult than real words then this would imply that the phonological assembly buffer is more impaired than the lexicons, as the person cannot rely on lexical information (Lesser and Perkins, 1999). The data in Case Study A also hint at the possibility of word retrieval difficulties. Again, this implies impairment at the level of the phonological assembly buffer and/or access to this. Whilst a more detailed assessment is required, it does appear that this patient is experiencing difficulties with the phonological assembly buffer.

Table 1. Modular account of selected disorders of language function (source: Lesser and Perkins, 1999)

Central to the modularity assumption is the notion of double dissociation. As Coltheart (2001:12) states, “With double dissociations we need two patients: patient A who is impaired on task X but normal on task Y, and patient B who is normal on task X but is impaired on task Y.” It is arguable that such so-called classic double dissociations cannot be found in brain-injured patients, in that it is rarely, if ever, the case that patients perform normally on any of the tasks. Double dissociations are, therefore, usually identified on the basis of relative impairment (see Crawford et al. (2003) for a discussion of fully operational definitions).

Modular accounts of language function have been challenged by connectionist models (Hinton and Shallice, 1991; Mozer and Behrmann, 1990). These differ from modular accounts in that they are computational models that learn by using algorithms to alter the weights of connections between so-called nodes. The representation of knowledge is distributed, with decisions arrived at by consensus of many simple calculations rather than any particular one. They therefore share the same processing resources. The relevance of connectionist modeling to neural implementation of cognitive processes is thought to be particularly strong when “the detailed pattern of breakdown and recovery of behavior in damaged connectionist networks resembles that of patients with cognitive impairments due to neurological damage” (Plaut, 1991:6). One such account is provided by Martin et al. (1996) who investigated recovery in deep dysphasia. They used a localist interactive model to simulate the spontaneous recovery of patient NC from deep dysphasia using Dell’s (1986) spreading activation model. From their observations of NC they hypothesized that (1) he had a fast rate of decay of activation of auditory-verbal short-term memory, and (2) this slowly returned to normal over time. They tested this hypothesis by manipulating decay rate and temporal interval between input and output and found that the model was capable of reproducing NC’s spontaneous recovery. This, therefore, goes some way to contraindicating the modular account of deep dysphasia in that activation was demonstrably interactive.

Symbolic v connectionist approaches

The polarization of modular versus connectionist models discussed in the preceding section may also be framed in terms of symbolic versus connectionist approaches, particularly in relation to language function. Historically cognitive psychologists have viewed the mind as a “symbol-manipulation machine, operating according to a well defined set of rules” (Cohen, 2000:249). The parallels with structural linguistics are apparent here, e.g. syntax is regarded as the grammatical rules for combining words (symbols) to create phrases and clauses. Symbolic accounts, therefore, suggest that language learning involves acquiring a system of symbolic rules. In contrast, connectionism offers an alternative non-symbolic account.

As already mentioned, double dissociations are frequently sought in order to test the modularity assumption. One candidate in the field of language learning is the differential performance in learning and using the English past tense in children with specific language impairment (SLI) and those with William’s Syndrome (WS). Children with SLI often have difficulty inflecting regular verbs for past tense (whilst demonstrating relatively normal production of irregular past tense verbs), whilst those with WS have difficulty inflecting irregular verbs for past tense (whilst demonstrating little difficulty inflecting regular verbs for past tense) (see Marchman, 1993:216). This apparent dissociation occurs without any known impairment in brain areas that are thought to be associated with language (e.g. Broca’s area, Wernike’s area).

Dual v single route accounts

The English past tense is an important test case in the debate between rule-based and connectionist accounts of human language processing (Rumelhart and McClelland, 1986; Pinker, 1991; Davis et al., 1996). This debate is often structured as a dual versus single route account. Dual route explanations of English past tense learning imply modularity. It is suggested that regular past tenses are generated by rules, whereas irregular past tenses are learned as exceptions, i.e. they are rote learnt and reproduced from associative memory. So, a symbolic rule process will generate regular past tenses by, for example, the addition of the suffix –ed to a word stem (e.g. walk → walk-ed; jump → jump-ed) and the irregular route associates a particular word stem to its stored irregular past tense form. Thus, different mental processes are seen as underpinning the generation of either regular or irregular past tenses. As Cohen (2000:251) points out:

“The regular route is symbolic because it is impervious to the phonological content of the particular word stem: It simply adds an /ed/ to whatever word it encounters. The irregular route, in contrast, needs to pay attention to the phonological content of the word stem in order to identify whether the current word is a candidate for irregular inflection. Rather than applying a rule, it associates that particular word stem with its irregular past tense form.”

It is well known that children produce overregularization errors when learning language, producing errors such as goed and drinked. Overregularizations are often viewed as evidence that children are indeed acquiring a system of learned rules: they are assumed to be the inappropriate application of the rule ‘add –ed to form a (regular) past tense’. This view is strengthened by the observation that children cannot be merely imitating and reproducing these forms, as they do not occur in adult language. Another well known aspect of learning past tenses is the so-called U-shaped learning profile. As McLeod et al. (1998:180) explain:

“…overregularization errors are surprising because they often occur after children have succeeded in producing the correct past tense of the verb. For example, children may correctly produce the past tense form went early in their third year and then produce goed at the beginning of their fourth.”

The dual route approach explains overregularization errors as an expected consequence of the transition from pure rote learning to partly symbolic rule-governed behavior. Reduction of errors occurs when “the representations of the irregular forms are sufficiently strong to resist interference from the rule-governed process” McLeod et al. (1998:180) (Figure 2).

[The dual-route model for the English past tense (based on Pinker and Prince 1988). The model involves a symbolic regular route that is insensitive to the phonological form of the stem and a route for exceptions that is capable of blocking the output from the regular route. Failure to block the regular route produces the correct output for regular verbs but results in overgeneralization errors for irregular verbs. Children must strengthen their representation of irregular past tense forms to promote correct blocking of the regular route.]

Figure 2. The dual route model for the English past tense (from McLeod et al. 1998:181)

In summary, dual route accounts assume that different resources are used in the processing of regular and irregular English past tenses. Under this scheme, therefore, selective impairments following brain damage (as evidenced by double dissociations) are readily interpreted in terms of the loss of particular processing components.

The major difference between dual route/symbolic accounts of English past tense learning and single route/connectionist accounts is, therefore, that the latter share the same processing resources when inflecting either regular or irregular verbs. Rumelhart and McClelland (1986) were the first to investigate children’s learning of inflectional morphology using a connectionist model. Their network was capable of explaining aspects of children’s learning of the past tense, including the U-shaped learning profile. Their work was, however, strongly challenged by Pinker and Prince (1988) who continued to argue for a symbolic account. Subsequent connectionist work has, however, begun to address some of the arguments they put forward. For example, Marchman (1993) demonstrated how a connectionist network (i.e. one in which regular and irregular verbs share the same single processing route) can give rise to asymmetric effects when the network is lesioned. She noted, amongst other findings, that regular verbs take longer to recover from damage than irregular verbs. This was explained by a variable contribution of the hidden units in the network to the processing of regular and irregular verbs. However, Marchman was unable to demonstrate a double dissociation in her model (the dissociations are always sparing of irregular relative to regular verbs). Work by Juola and Plunkett (1998) did, however, identify double dissociations between regular and irregular verbs in connectionist networks. The conclusion was similarly that the processing of regular and irregular words is not divided equally amongst the network connections.

Studies by Plunkett and Marchman (1993, 1996) demonstrated that a single route mechanism was capable of producing a developmental profile similar to the overregularization errors in the speech of English children reported by Marcus et al. (1992) (Figure 3).

[A comparison of the overregularization errors of Adam, a child studied by Marcus et al. (1992), and those produced by the Plunkett and Marchman (1993) simulation. The thin line shows the proportion of errors as a function of age (Adam) or vocabulary size (simulation). The thick lines indicate the percentage of regular verbs in the child’s/network’s vocabulary at various points in learning. ]

Figure 3. A comparison of overregularization errors (from McLeod et al. 1998:186)

Conclusion

Both modular and connectionist accounts of the cognitive processes underpinning language function seem capable of explaining the performance of normally developing children (here we have focused on the acquisition of regular and irregular past tenses) and the effects of brain damage and rehabilitation in adult aphasics.

Much of cognitive psychology assumes that the cognitive system is composed of a collection of encapsulated processing modules, each dedicated to performing a particular cognitive function. As we have noted, selective impairments of cognitive tasks following brain damage, are readily interpreted in terms of the loss of particular processing components. Such an approach leads to the definition of specific disorders such as word-sound deafness.

Connectionist accounts propose plausible alternative explanations. However, one needs to be cautious in interpreting the findings. For example, just because a connectionist network can perform a particular task does not prove that children use the same mechanism, only that they might do so. However, such models demonstrate that it is not essential to invoke a separate, symbolic system to explain the processing of verbs. Instead, it is possible that both regular and irregular forms are processed in the same manner and by the same mechanism.

We have also seen how connectionist models can replicate the pattern of children’s development. However, it is not clear whether the learning mechanisms are the same, or even similar, in child and machine.

References

Bauer, R.M. (1993) ‘Agnosia’ in Heilman, K.N. and Valenstein, E. (eds) Clinical Neuropsychology New York: Oxford University Press.

Cohen, G. (2000) ‘Studies in language processes’ in Cohen, G., Johnston, R.A. and Plunkett, K. (eds) Exploring Cognition: Damaged Brains and Neural Networks Hove: Psychology Press.

Coltheart, M. (2001) ‘Assumptions and methods in cognitive neuropsychology’ in Rapp, B. (ed) The Handbook of Cognitive Neuropsychology Philadelphia: Psychology Press.

Crawford, J.R., Garthwaite, P.H. and Gray, C.D. (2003) ‘Wanted: fully operational definitions of dissociations in single-case studies’ Cortex 39, 357-370.

Davis, M. H., Marslen-Wilson W. D. and Hare, M. (1996) ‘Representing regularity: the English past tense’ in Cottrell, G.W. (ed) Proceedings of the Eighteenth Annual Conference of the Cognitive Science Society Mahwah, NJ: Lawrence Erlbaum Associates.

Dell, G.S. (1986) ‘A spreading activation theory of retrieval in sentence production’ Psychological Review 93, 283-321.

Ellis, A.W. and Young, A.W. (1988) Human Cognitive Neuropsychology London: Lawrence Erlbaum Associates.

Franklin, S. (1989) ‘Dissociations in auditory word comprehension: evidence from nine fluent aphasic patients’ Aphasiology 3, 189-207.

Franklin, S., Turner, J., Lambon Ralph, M.A., Morris, J. and Bailey, P.J. (1996) ‘A distinctive case of word meaning deafness?’ Cognitive Neuropsychology 13, 1139-1162.

Hinton, G. and Shallice, T. (1991) ‘Lesioning an attractor network: investigations of acquired dyslexia’ Psychological Review 98, 74-95.

Juola, P. and Plunkett, K. (1998) ‘Why double dissociations don’t mean much’ Proceedings of the 20th Annual Conference of the Cognitive Science Society (CogSci-98) 1-6.

Lesser, R. and Perkins, L. (1999) Cognitive Neuropsychology and Conversation Analysis London: Whurr Publishers.

Levelt, W.J.M., Schriefers, H., Vorberg, D., Meyer, A.S., Pechmann, T. and Havinga, J. (1991) ‘The time course of lexical access in speech production: a study of picture naming’ Psychological Review 98, 122-142.

Marchman, V.A. (1993) ‘Comnstraints on plasticity in a connectionist model of the English past tense’ Journal of Cognitive Neuroscience 5, 215-234.

Marcus, G., Ullman, M., Pinker, S., Hollander, M., Rosen, T.J. and Xu, F. (1992) ‘Overregularization in language acquisition’ Monographs of the Society for Research in Child Development 57.

Martin, N., Saffran, E.M. and Dell, G.S. (1996) ‘Recovery in deep dysphasia: evidence for a relation between auditory-verbal STM capacity and lexical errors in repetition’ Brain and Language 52, 83-113.

McLeod, P., Plunkett, K. and Rolls, E.T. (1998) Introduction to Connectionist Modelling of Cognitive Processes Oxford: Oxford University Press.

Mozer, M. and Behrmann, M. (1990) ‘On the interaction of selective attention and lexical knowledge: a connectionist account of neglect dyslexia’ Journal of Cognitive Neuroscience 2, 96-103.

Parkin, A.J. (1996) Explorations in Cognitive Neuropsychology Hove: Psychology Press.

Pinker, S. (1991) ‘Rules of language’ Science 253, 530-535.

Pinker, S. and Prince, A. (1988) ‘On language and connectionism: analysis of a parallel distributed processing model of language acquisition’ Cognition 28, 73-193.

Plaut, D.C. (1991) Connectionist Neuropsychology: The Breakdown and Recovery of Behavior in Lesioned Attractor Networks PhD Thesis, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA.

Plunkett, K. and Marchman, V. (1993) ‘From rote learning to system building: acquiring verb morphology in children and connectionist nets’ Cognition 48, 21-69.

Plunkett, K. and Marchman, V. (1996) ‘Learning from a connectionist model of the acquisition of the English past tense’ Cognition 61, 299-308.

Rumelhart, D.E. and McClelland, J.L. (1986) ‘On learning the past tense in English verbs’ in McClelland, J.L., Rumelhart, D.E. and PDP Research Group (eds) Parallel Distributed Processing : Volume 2 Cambridge, MA: MIT Press.

Takahashi, N., Kawamura, M., Shinotou, H., Hirayama, K., Kaga, K. and Shindo, M. (1992) ‘Pure word deafness due to left hemisphere damage’ Cortex 28, 295-303.

Appendix 1: Case Study A

History

A 23 year old right-handed woman suddenly developed a severe headache whilst doing aerobics. It was the severest headache that she had ever experienced; she fell to the ground clutching her head. An ambulance was called and she was taken to Accident and Emergency. On arrival 30 minutes after the onset of the headache, she appeared confused; not apparently understanding what was said to her. However, her spoken speech was relatively normal, although she occasionally used an incorrect grammatical construction and nonsense words, e.g. she referred to the doctor as ‘dister’. She had developed a right-sided weakness involving the face and limbs and had severe neck stiffness, sensation could not be assessed. Her blood pressure was normal. A provisional diagnosis of cerebral hemorrhage was made.

Medical investigations and interventions

An urgent CT brain scan confirmed a large hemorrhage in the left hemisphere in the middle cerebral artery territory, i.e. parietal and posterior frontal areas, with displacement of the brain to the right. In addition, there was some blood in the ventricles, i.e. subarachnoid hemorrhage. An urgent angiogram via arterial catheter demonstrated an aneurysm of the middle cerebral artery with extensive spasm of all the left-sided arteries. A calcium antagonist was given (nimodipine) in an attempt to protect the brain from arterial spasm.

Twelve hours later the clinical state remained the same. It was decided to evacuate the clot and clip the aneurysm. This was successfully achieved. Immediately postoperatively she was mute with a dense hemiplegia but over the next three weeks the hemiplegia improved and her spoken speech returned. At this time she seemed to understand very little, e.g. when asked to close her eyes she looked perplexed and did not perform the task. Over the next six weeks her condition improved and she was well enough to undergo neuropsychological assessment.

Neuropsychological assessment

Her speech was fluent with occasional grammatical errors when shown pictures of actions, e.g. ‘the girls is skipping.’ She also made errors in naming common objects, e.g. items of clothing. There was particular difficulty in retrieving proper names, e.g. naming a picture of Tony Blair or saying who the President of the USA was. Her most striking deficit however was in repetition. Her digit span was 1. She was unable to do simple addition or subtraction and was poor at multiplication.

Further neuroimaging investigations

An MRI of the brain demonstrated patchy infarction centered upon the posterior parietal lobe, involving the angular gyrus and parts of the superior temporal gyrus but Wernicke’s area and the frontal lobes were intact. A repeat angiogram showed complete occlusion of the aneurysm and the arterial spasm had resolved

Conclusion

This case has elements three types of aphasia:

• conduction aphasia, i.e. poor repetition with fluent speech and reasonable comprehension
• Broca’s aphasia, i.e. non-fluent speech, poor repetition and good comprehension
• Wernicke’s aphasia, i.e. fluent speech, fair repetition and poor comprehension

Anatomically, conduction aphasia is associated with lesions in the posterior parietal area (supra-marginal and angular gyrus) and this localized area could lay claim to being the anatomical area for a distinct module of cognitive function. It is important however to remember that there are large fiber tracts passing beneath this area connecting occipital and parietal areas with temporal regions.