6.3 Psycholinguistic considerations
The need for a heterogeneous theory of morphology has also been the subject of intense dispute in the
psycholinguistics and neurolinguistics community. The genesis of the dispute was the claim by David Rumelhart and
James McClelland (1986b) that a connectionist network that performs pattern association could learn the past tense
syste mof English in roughly the way a child does. Steven Pinker and colleagues, in part drawing on work by many
others (Pinker 1999); Pinker and Prince 1988); 1991; Clahsen 1999 and references therein)have amassed considerable
evidence against this account, arguing that something more than pattern association must be taking place. First
consider what the past tense network does. Ostensibly it learns to perfor mthe task“Given the verb X, for mits past
tense.”In fact it actually learns a different task:“Given a sequence of phonemes, produce the sequence of phonemes
stipulatedto be associated with it.”That is, as Pinker emphasizes, thenetworkmodel has no notion ofword. Of course
it doesn't have to, because it never has to deal with sequences of words, only with words given in isolation. However,
because it identifies phoneme sequences rather than words, it cannotdistinguish theverbring(withpastrang) fromring
(ringed)andwring (wrung),orhang (hung)fromhang (hanged).
In addition, the network also has no notion of verb or of past tense. This is because the network is not intended as a
“general morphology machine”: it doesn't have to tell verbs fro mnouns, so as to be able to decide whether to apply
past tense or plural, for example. Nor does it have to decide whether the appropriate ending on a verb in a particular
context is past tense or, say, present participle—because there are no contexts. Part of the overall connectionist
philosophy (e.g. Rumelhart and McClelland 1986a; Elman et al. 1996) is that discrete symbols such as Verb and Tense
are rendered unnecessary by the architecture of networks. But in this case the only reason they appear unnecessary is
that virtually everything of linguistic relevance has been bled out of the task the network has been designed to
perform.^81 It is therefore hard to take seriously the manifold claims, still being repeated (e.g. Seidenberg and
MacDonald 1999);
LEXICAL STORAGE VS. ONLINE CONSTRUCTION 163
(^81) Plunkett and Juola (1999) develop a network that is intended to add both plural to nouns and past tense to verbs, so it appears to be on its way to being a“general
morphology machine.”However, theauthors note(484) that“thenetwork,bydesign,only has accesstophonologicalinformation.”Thus ithas nowayof knowingwhether
the phonological sequence /saedl/ is meant to be the noun or the verbsaddle, hence how it should be inflected. And the authors admit (484):“Unfortunately, it is not
possible in the present simulations to model this process.”They then briefly describe a modification that adds random“pseudo-semantics,”whichis supposed to simu-late
the distinction between the noun and verb uses ofsaddle, for instance. This hardly models the noun-verb distinction, though; it could just as well simulate the difference
betweenring andwring.