The Generation of Continuous Semantic Constraints from Semantic Propositions

Language comprehension is an exceedingly complex process which requires the extensive use of many di f ferent kinds of information in order to be successfully accomplished. One potent ial ly very important type of information which has to date been largely ignored is the degree to which possible interpretations are sensible. While the sensibleness of candidate interpretations has long been recognized to be important, sensibleness has usually been treated as if it were an all-or-none property. However, it is clear that many things are more-or-less sensible and, therefore, the relat ive sensibleness of alternative interpretations may well be extremely useful information. For example, Oden (1977) has argued that degree of sensibleness information is required in order to disambiguate sentences to obtain the meaning that people normally do and has proposed language processing mechanisms which would use this information. The degree of sensibleness of an interpretat ion depends on the degree to which the semantic constraints of that interpretat ion are sat is f ied. Therefore, to account for the continuous nature of sensibleness, semantic constraints must be fuzzy restr ict ions (Zadeh, 1975). A semantic constraint w i l l be defined to be a function associated with a part icular semantic relat ion which specif ies, for every combination of semantic elements which may enter into that re la t ion, the degree of sensibleness of the resultant semantic structure. The present paper outlines how such semantic constraints may be generated from the kinds of knowledge already represented in current semantic memory models (e .g . , Norman & Rumelhart, 1975), plus the fuzzy predicates and operations which w i l l be necessary in order to handle other problems l ike the continuousness of subjective class membership. Defining semantic constraints to be functions makes it natural to think of complex semant i c constraints as being compositions of simpler constraints. Furthermore, since semantic constraints are considered to be bound to part icular semantic relat ions, the decomposition of a constra int may be expected to paral lel the decompos i t ion of i t s associated semantic re lat ion. This appears to be what happens in most cases, but there are certain "configural constraints" which do not seem to be derivable from component constraints corresponding to pr imit ive semantic relations (see Oden, 1977, for deta i ls ) . However, such configural constraints seem to be re lat ive ly exceptional and, consequently, semantic constraints w i l l s t i l l be "cognit ively economical." More importantly, it w i l l be argued below that elementary (non-composed) semantic constraints, whether d i rect ly associated with pr imit ive semant i c relations or conf igural, are based upon specific semantic propositions which would be in semantic memory anyway. In a fundamental sense, a l l knowledge is constraining. For example, knowing that it is -15° today affects the sensibleness of the statement "Maxine went swimming in Lake Mendota th is morning." More generally useful knowledge specif ies information about the normal and/or necessary properties of things which may enter into pa r t i cular case relations with part icular verbs. The most elementary knowledge of th is sort (select ional restr ic t ions) is often considered to be part of the basic meaning of the verb. However, the more interesting semantic constraints are those which are based on much less elementary knowledge, such as that only people normally drive trucks, which we might represent as: