McClelland and Seidenberg2000), that theneural networkapproach as presentlyconceivedpotentiallyyields thekey to
the full productivity of language.
Pinker and colleagues also adduce experimental evidence against a pure network account, drawing on the statistical
distribution of regulars and irregulars, on speakers' judgments and reaction times in producing and recognizing past
tenses of known and novel verbs (Bybee and Moder 1983; Prasada and Pinker1993), and on the behavior of
agrammatics and individuals with Specific Language Impairment (Gopnik 1999; van der Lely1999) and with Williams
Syndrome (Clahsen and Almazan 1998). They show that the pattern association account is satisfactory for irregular
verbs (here, semiproductive patterns), and even provides considerable insight. But they also show that the formation
of regular past tenses is a phenomenon of quite a different character, not captured by the Rumelhart and McClelland
model nor by subsequent connectionist models in a similar spirit such as those of Plunkett and Marchman (1991) and
Hare et al. (1995). The experimental results have been replicated with other morphological systems in which a regular
default pattern is overlaid by a syste mof se miproductive patterns: Ger man noun plurals (Marcus et al. 1995; Clahsen
1999), Hebrew noun plurals (Berent et al. 1999), and Japanese noun-forming suffixes (Hagiwara et al. 1999).
I acknowledge that there is a great deal of controversy surrounding these results (see for instance many of the
commentaries accompanying Clahsen 1999). But a much more important point has become lost in the debate. Being
able to learn to produce past tenses on demand is not enough for learninglanguage. In particular, Marcus (1998, 2001)
argues that a multilayer network trained by back-propagation—the standard network used in these models—is in
principle incapable of extracting the sort of regularity necessary to account for productive morphology. The reason is
thatsucha regularitymust beformulated interms ofavariable:“thisgeneralization applies toanythingI encounter that
belongs to such-and-such a category, regardless of its other features.”As observed in section 3.5, Marcus shows that
pattern associators simply cannot encode variables.
Pinker characterizes the productive process of past tense formation as a rule with a variable:“To for mthe past tense
ofany verb, add -d.”Pinker's clai mis that a network instantiating pattern association cannot learnrules. The present
account is slightly different: regular past tenses are formed by the free combination of the lexical entry (5) with a verb
stem, in accordance with the affix's structural requirements. Instead of Pinker's special rule or a somewhat more
general rule“Add affixes to stems,”there is the maximally simple and general ruleCOMBINEorUNIFY. Under this
interpretation, though, Marcus's argument