At the moment in time marked by *, the question faced by the language processor is whetheronis going to designate
where the appleisor where it is to beput. It turns out that at this point, subjects already start scanning the relevant
locations in thearray in order to disambiguate thesentence (Is theremore than oneapple? Is therean applealready on
the towel?). Hence visual feedback is being used to constrain syntactic structure early on in processing.
The present architecturemakes clear howthiscomes about: thesyntax–semanticsinterfacelinkscompeting analyses in
syntax with competing conceptual structures. In turn, these alternatives are picked up by an interface that connects
conceptual structure to the visual syste m(Macna mara 1978; Jackendoff 1987; Landau and Jackendoff 1993; see also
Chapter 11). Visual search is used to “clamp” the contextually appropriate interpretation; inhibition of the
inappropriate interpretations is passed back through the interfaces to conceptual structure and then to syntax. This
case thus presents a parallel to (2), with the interactions just moved one component over in the system: syntax via
conceptual structure to vision instead of phonology to syntax to conceptual structure.
Similar considerations arise in production. For instance, Levelt's (1999)“blueprint of the speaker”(p. 87) includes a
component of“self-perception,”in which the speaker monitors phonological structure internally (“hears what he/she
is about to say”) and uses this to revise or repair the conceptual structure being encoded as a sentence. Levelt's
diagra mtreats this co mponent as a direct feedback fro mphonological structure to conceptual structure. But this is
oversimplified, as Leveltnotes (p. 88). Phonologicalstructure can affectsemantics only through the rules of grammar;
thus“what I a mabout to say”has to be interpreted via the phonology–syntax and syntax–semantics interfaces. In
other words, this is the inverse of the situation illustrated by (2) for language perception.
The symmetry of this situation is reminiscent of Garrett's (2000) proposal that the production processors are used in
perception and vice versa: the routes used for feedback in one are used for“feedforward”in the other. The present
approach explains whythisproposal should make sense: theinterfaceconstraintsdon'tcare inwhichdirectiontheyare
applied.
To su mup the argu mentso far, se mantics cannotdirectly influence phonologicalanalysis—or vice versa^98 —and visual
context cannot directly influence
IMPLICATIONS FOR PROCESSING 203
(^98) Careful readers may note that Fig. 7.2 includes a direct phonology-semantics interface, and may wonder if this might be used for the feedback in these examples. The
answer is no: this inter-face is not sufficient to mediate completelybetween the two levels. We have seen (Ch. 5) that it includes lexical structure (e.g. /kset/ means CAT)
and the relationof intonationto focus; Ch. 8 will suggest that it also includes some weak constraints between linear order and meaning. But this interface does notinclude
anything about such crucial aspects of meaning as argument structure and modification. These necessarily invoke the interfaces with syntax.