PREDICTING SYNTAX FROM SEMANTICS 501
statements in the lexicon rather than in the syntax, e.g. redundancy rules,
relation-changing rules such as passive, and control specifications for verbs.
Other theories, such as Government-Binding theory [GB] (Chomsky 1981,
1986) and RRG fall somewhere between these two extremes, as they make
use of both lexical and syntactic rules and principles.
As to the question of whether the lexicon is used as a source for the
prediction of at least some syntactic facts, there are two schools of thought.
In the first school the lexical entries of predicates merely list the syntactic
subcategorization frames, the prepositions, and the complements that the
predicates occur with and hence do not use the lexicon as a source for the
prediction of syntactic facts but use it more as an ad hoc tally sheet.
Theories that are of this type include HPSG, LFG, and that of Jackendoff
(1972, 1976, 1983). Radford (1981), in his discussion of theories of control,
pinpoints the crucial weakness of this approach:
Firstly, arbitrary lists of properties associated with predicates have no pre
dictive or explanatory value: ask the question "How do you know this is a
verb of subject control?", and you get the non-answer "Because it's listed
as a verb of subject control in the lexicon." Secondly, treating control...as
a lexically governed phenomenon implies that control properties are
entirely arbitrary, and hence will vary in random fashion from dialect to
dialect, or language to language: this would lead us to expect that the
counterpart of [e.g. John persuaded Bill to leave] in some other dialect or
language would have subject control rather than nonsubject control...But
as far as we know, this is not the case (p. 381).
Thus simply listing the complement types that a predicate takes in its lexical
entry (in either syntactic or semantic terms or both) results in an arbitrary
list and fails to capture any generalizations about complement selection.
In the second school of thought attempts are made to derive syntactic
facts from lexical representations in a principled way. The best known
example is GB with its Projection Principle. Chomsky (1986:86-90) has
gone beyond the Projection Principle to suggest that the syntactic sub-
categorization features of lexical items, in particular verbs, can be predicted
from the aspects of their lexical representation relating to their argument
structure. Using the example of persuade, he argues that the relevant syn
tactic subcategorization features follow from the "s[emantic]-selection"
properties of a verb: persuade takes two arguments, a goal and a proposi
tion, and the "canonical structural realization" of a goal is an NP and that
of a proposition is either an S or an NP with a propositional interpretation.
He concludes that "[i]f the argument can be generalized..., we may con-
