Reentrancy in feature structures

[Carl Pollard]

Hi everyone,

John Beavers said:

>
Well, it seems to me that you've just replaced co-identification with some
notion of equivalence, and I don't think this makes a difference.  It
certainly wouldn't make any difference in whatever way you build such
models (even the problem of infinite recursion is still an issue if a path
is constrained to be equivalent to a superpath).  As far as the models
themselves, well, any dag you get as a model can be converted into a tree
algorithmically but you lose the equivalence information as part of the
data struture.  It seems to me that's the only difference, and depending
on whether you find that kind of information useful it could be an
unimportant difference.
>>

The last sentence is the key question: whether there is a NEED to
distinguish identity of the nodes at the end of two paths from
structural isomorphism between the two substructures rooted at
those nodes. The main place this difference has beem used in HPSG
has been distinguishing between coindexing (two paths leading to
the same substructure of type _index_) and cases where two NPs
have distinct but structurally isomorphic indices (which in the
feature geometry of Pollard and Sag 1994 means that the two NPs have
the same agreement features. If one wants to give this up, fine,
but then binding (in the sense of Principle A and Principle B effects)
the link between (what has been called) coindexing and agreement
must be taken care some other way.

Carl