mightfit, it is initially integrated“promiscuously.”Likewise, if a position in a structure can be identified with
more than one possible item, they are initially all integrated with the position, in competition with one
another.- Resolutionis the collection of processes by which multiple possibilities for integration are pared down. These
come in at least twosubvarieties. First, the integrative processors must check for full integration of the input.
For instance,the phonology integrativeprocessor rulesout thechoiceofaplusapparentinexample(2)above,
because thesamesegmentisassignedtotwowordsatonce.Second,ifa structureatonelevelisincoherentor
disfavored by its integrative processor (i.e. loses out to competitors in the integration process), the interface
processors cansend inhibitionup-ordownstreamtotheassociatedstructures atother levels. Theoverallgoal
of resolution is to produce a maximally favorable single structure at all levels.
Itshould also be remembered that eachof these subprocesses is realized neurally. Thus theyare notinstantaneous, but
rather havea timecourse—a risetimebeforemaximum intensityisachieved;furthermore,themaximumintensitymay
differ fro mite mto ite m. It is here, for instance, that it makes sense to localize the fact that reaction ti me in lexical
decision tasks depends on word frequency (Oldfield and Wingfield 1965).
7.4.3 Lexical access in production
Because most work on language perception is concerned with syntactic parsing, theories of perception cannot help
dealing with combinatoriality. However, a great deal of the production literature (e.g. Caramazza and Miozzo 1997;
Engelkamp and Rummer 1998) seems to havefocused on relativelynon-combinatorial tasks like picturenaming. They
have therefore been able to conceptualize production simply in terms of spreading activation in a static network: a
visual stimulus drives one end of the network, and the phonology of a word comes out the other end. Such a
conception, however, is simply not up to the task of producing real language, since its machinery leaves no room for
actually building sentences.
A more sophisticated tradition (e.g. Levelt 1989; 1999;Roelofs 1997; Levelt et al. 1999; Dell, Burger, and Svec 1997)
recognizes thecriticalimportanceofcombinatoriality. Lexicalaccess is stilltreatedin terms ofspreading activation,but
when it comes to building a sentence, the theory appeals to a more algorithmic view in which linguistic items are
integrated into a skeleton of slots. Using a stricter distinction between long-ter mand working me mory, and the
distinction among the four subprocesses above, we can perhaps refine this view a little.
Suppose the conceptual department of working memory contains some thought that the speaker wishes to express.
(Along with Levelt, I will not concern