Default unification: online vs. offline

[Luis Casillas]

I was just wondering if I could get any thoughts from people on the
motivations behind the restriction of default unification in HPSG to the
type hierarchy, i.e. the assumption that all maximally specific types
specify only "hard" constraints and no default information.

The kinds of argument I've heard in favor of it are generally of the
kind "more defaults mean more computational complexity" (or alternately,
"a much more complicated and less understood feature logic missing
tons of very nice properties"). Essentially, that the type hierarchy
is a finite object, where we can feasible compile out the effects of
default constraints at the bottom level and thus get away with defaults
without suffering the extra complexity, but on the other hand the set of
objects characterized by the grammar is infinite, and by allowing online
defaults we pay a big cost in complexity.

However, am I missing other arguments in favor of this restriction
other than the computational one? Because, frankly, if that is the
one argument, it sounds a bit too much like letting our interest in
computational applications of the theory drive its formal devices,
rather than letting the actual linguistic and psychological data tell us
what they should look like.

--
Luis Casillas
Department of Linguistics
Stanford University