The discussion of current approaches for parsing and generation can be summarized as follows: parsing and generation, to be goal-directed, differ basically with respect to the order in which the literals of the body of a clause are selected. For parsing, for example, [Shieber1988, Gerdemann1991] have used the leftmost selection strategy, where for generation [Shieber et al. 1989, Gerdemann1991] use the semantic-head first selection function. The latter should be seen more precisely as a ``preference-based'' selection function, since in the case a rule has no semantic head, the leftmost element is chosen, or if two elements share the semantics with the mother node, the left one is selected.
However, it is very easy to combine these different strategies used in
parsing and generation, such that the selection function expresses a
preference for goals with certain features instantiated.
Since we want to obtain an input driven algorithm, the essential feature for
parsing should be the PHON path (more precisely the path is
,
that is the path to the list value of the difference list) and for
generation it should be the SEM path. We will call this certain
feature the essential feature . Then the selection function
can be defined such that it selects the leftmost element from the
body whose essential feature is instantiated, i.e., whose exists
and is a non variable value. If such an element does not exist it chooses
the leftmost element. Now, if we abstract away from a concrete
essential feature by assuming that is a variable, then we can
define this selection function more formally as follows:
Now, in order to use this selection function for parsing or generation we have to specify a path that defines the essential feature (i.e., the phonological or semantic path). Since, the value of this feature will be a string or semantic expression, this means that the selection function prefers those goals which are instantiated with a string or semantic expression. However, now, the grammar itself will be an important source of control, since it defines how information is decomposed (or composed depending on the point of view) in the rules. For example, if the phonological information is expressed as difference lists and partial strings are combined by string concatenation then the selection function ``realises'' a leftmost strategy. Similarly, if all rules define a semantic head relation simulates the semantic head first relation. These can both be true at the same time. Moreover, if the grammar rules are attached with some preference values, the selection function can very easily be adapted to take into account such preference information. This would help to achieve a more careful selection.