Fig. 7.2 The parallel grammar implemented as a processing architecture
acoustical information—a frequency analysis of a continuously varying speech signal—into a discretely
segmented phonetic structure. Another such process uses syntactic parses to construct specifications of
semantic roles. Let us call such processesinterfaceprocesses.^97 For each set of interface constraints that links a
pair of levels, the processor needs an interface processor thataccomplishes such linking—or perhaps itneeds
one in each direction.- In addition to these two types, the f-mind needs processes that take as input full or partial structures in a
particular format in working memory, and relate them to or construct new structures in the same format.
Classiccases are rulesofinference, whichderivenewconceptual structures fro mexisting ones, and principles
ofmental rotation, whichderivenewvisual/spatial structures fromexistingones. Anothersuchcasemight be
checking to see whether two words rhyme, which compares two phonological structures. I'll call these
inferentialprocesses.
Given these distinctions, it is possible to describe the logic of processing in terms isomorphic to the rule types in the
parallel grammar. Fig. 7.2 illustrates.
IMPLICATIONS FOR PROCESSING 199
(^97) In Jackendoff (1987) these were called“translation”processes. The reason for the change in terminology is that these conversions do not preserve“meaning”—in two
respects. First, we are abandoning the view that f-mental structureshave meaning (Ch. 2) , that is, they are not to be viewed asrepresenting anything. As far as the brain is
concerned they are just pure structure. Second, the relations that an interface implements between structures do not preserve structure in the way the term“translation”
would suggest. As seen in Chs. 1 and 9 5 and section 7.5, it is better to think of interfaces as implementing only partial homologies between two forms of structure.