Van Noord and Bouma propose to treat lexical rules as constraints on lexical categories, and to use delayed evaluation techniques. These constraints are implemented as rules which must be satisfied by the lexical entry of a word in a particular category. The constraints are evaluated with respect to the base (or ``stem'') form of a word in the lexicon. The true lexical entry for the word used in an attempted parse results from evaluation of constraints with respect to the base form.
The delayed evaluation techniques prevent constraints from being evaluated until enough information is available to do so. This means that constraints may actually only be partially evaluated at any step in the application of multiple constraints to a single lexical entry. The benefit of these techniques is that parsing mechanisms can interact with lexical information, allowing constraints from both structural and lexical levels to apply simultaneously as input is processed. This essentially means that sentence-level semantics can be encapsulated in a lexical rule. This accommodates a notion of a construction (e.g. Goldberg goldberg:95) in which a particular syntactic form is associated with a specific interpretation, and provides an elegant way of defining precisely how different parts of a sentence combine to produce an interpretation. We will see additional need for such constructions in Chapter 4, and there is other work in HPSG (Sag in press) which argues for the specification of non-compositional phrasal-level semantics.
The van Noord and Bouma approach accommodates most of the characteristics of adjuncts well. A single lexical entry is necessary for each adjunct, and they allow for both restrictive and operator adjuncts by requiring the appropriate semantic combinations to be specified in the mod field of the adjunct, following Kasper's (1993) approach. Mittelfeld phenomena are handled by allowing for the insertion of the adjuncts at any point in the verbal subcat list.
It is noted by van Noord and Bouma that their approach is flexible enough to accommodate various approaches to the ordering of adverbials on the subcat list. Although the lexical constraint controlling the addition of adjuncts as defined in their paper assumes that the adjuncts are inserted into the subcat list in order of semantic obliqueness (adopting Kasper's idea of semantic combination from narrow to wide scope), there is nothing in their methodology which restricts the definition of the constraint. It is difficult, as mentioned above, to see precisely how syntactic ordering effects could be accommodated in an approach that relies entirely on semantic obliqueness. Changes in the ordering on the subcat list, however, would require radical changes in van Noord and Bouma's definition of the constraint adding adjuncts. In particular, if the ordering on the subcat list were changed to reflect surface order of the adverbials, their recursive approach to semantic composition would no longer suffice. Other mechanisms, analogous to the linear precedence (LP) constraints which are required to handle word order restrictions in their existing approach, would be necessary to control semantic composition. These mechanisms could only be applied to a fully expanded subcat list and thus would prevent semantic content from being truly recursively computed.
Redundancy constraints are a problem in the vannoord_bouma:94 approach, as in all other approaches discussed here. They could conceivably be defined in the requirements in the mod field specifications, but again it is not clear how to do this in a straightforward manner.