against the network approach still goes through. The rule has two variables: “Combine anything with anything
appropriate”; and the lexical ite m-ddefines what is appropriate using a variable:“I combine withany verb.”So the
present account poses the same principled difficulties for a network of the usual sort.
The present approachin factmakes clearer whatis at stake in thisargument.As stressed in Chapter3), theprincipleof
free combination is what makes language what it is. Connectionist approaches do give us considerable insight into
semiproductivity. But if a multi-layer network trained by back-propagation in principle cannot account for something
as combinatoriallytrivial as the regular past tense, there is no hope of scalingup current connectionist solutions to the
past tense to the rest of language, where free combination reigns supreme. Rather, as suggested in section 3.5, the
challenge to network approaches is to develop a robust device for encoding variables—not to continue to try to live
without them. Accepting this challenge could, I believe, lead to far more meaningful dialogue between network
modelers and theoretical linguists.
To my knowledge it has not been observed that a heterogeneous morphological theory presents an interesting
challenge to acquisition. The child, after all, does not have a“Maxwell's demon”that miraculously sorts the words in
the environment into regular and semi-regular cases. The same morphological relationship (say causative formation)
may be productivein onelanguage (e.g. Turkish) but only semiproductivein another (e.g. English). Nor does relatively
high frequency signal a regular pattern. For instance, German regular plural nouns are considerably less frequent than
semiproductive plural forms in speech, counting either tokens or types (Marcus et al. 1995). My inclination is to think
that the child's brain takes a catholic approach, trying to analyze new morphological patterns both ways at once, and
that eventually one mechanism extinguishes the other. We return to this issue in section 6.9.
6.4 The status of lexical redundancy rules
If the irregularateis a separate lexical ite mfro meat, then how are they related? As just mentioned, most of the
experimental work in Pinker's traditionviews the task of relating them as“Giventhe verbeat, for mits past tense”; the
connectionist tradition treats it similarly but in less structured terms. However, performance on this experimental task
is only a reflectionofdeeperlinguisticf-knowledge thatis normallyemployed intasks more relevant toactuallanguage
use. For instance, upon hearingatein a sentence, one does not look in the lexicon for some verb that this is the past
tenseof,“undoing”thephonological“change”fro m/iy/to/ey/. Ratheronelooks up thephonologicalfor m/eyt/ in