Italian, gender agreement endings on adjectives; these have no counterpart in the conceptual structure to be
expressed.^103
A similar but more complex story is told in perhaps the most detailed and influential tradition in speech production,
that of Levelt and colleagues, referred to above. Levelt et al. argue, on the basis of substantial experimental evidence,
that lexical access proceeds in two stages, selection of lemmas and selection of word forms or lexemes. Levelt (1989)
characterizes the lemma (following Kempen and Hoenkamp 1987) as the complex of a word's semantic and syntactic
features; in some later work it is characterized instead as the syntactic features alone. The initial process of lexical
access is argued to involve only lemma selection;no phonological forms have yetbeenactivated. After conceptual and
syntactic integration and resolution have been achieved,there is a second call to long-term memory to retrievea word
form (or lexeme) that provides the phonology associated with the lemma's syntactically integrated form. Since the
syntacticallyintegrated for mincludes gra m matical inflectionand thelike,wecansay thatgoandwentare differentword
forms associated withthesame lemma GO, and thatgrammaticalitanddoare wordforms without associated lemmas.
This is not the place to go through the extensive experimental evidence adduced in favor of this view—and against it
(e.g. Dell1986; Dell,Schwartz etal. 1997).However, twothingsmakemeuneasyaboutitontheoreticalgrounds. First,
recall the logic of priming in speech perception. Phonological activation ofbughas to result in semantic activation of
bugin long-term memory, which in turn activatesthe semantics ofinsect, whichactivatesthe phonology ofinsect. If the
lexicon is divided into lemmas and word forms, then the claim must be that lemmas can activate word forms in
perception, but not in production. There may indeed be such an asymmetry, but it does not follow from anything.^104
IMPLICATIONS FOR PROCESSING 213
(^103) Caramazza and Miozzo (1997) champion a lexical network in which semantic features activate phonology, which in turn activates syntactic features. They fail to consider
the need for syntactic integration and for an independent working memory; theirtheory is based on a picturenaming paradigm, where syntactic integration is unnecessary.
The sorts of phonological choice mentioned here, which depend on syntactic integration, are therefore completely unavailable to their model.
(^104) Peopleoccasionally talk about havingtwodifferent lexicons,one for perceptionand one for production. Ifind this odd, giventhattheyinvolvethesame kinds of structure,
and given that perceived words can prime production. I tend to think that the differences between the two tasks are due to different processing problems rather than
different lexical data bases. For one thing, perception is recognition memory for lexical items, while production is the much more difficult recall memory. Production also
requires the assemblyand executionof a motor program, adding to its difficulty. Finally, access to the lexiconthrough semantics in productionmay be inherentlyless stable
than access through phonetics in perception—production has no external signal to which processing can be“clamped.”