trees

[Carl Pollard]

Hi Andrew,

Much of what you are asking about is SWB-specific (I think), and for
them to respond to. As I understand Tibor, he would like to scrap a
lot of the machinery of Pollard and Sag 1994 and instead take GPSG as
a starting point, adding conservatively to it. In terms of the
Pullum-Scholz GES-MTS dichotomy, PS94 is pure MTS: there is a universe
of candidate structures, the grammar is a logical theory, the
candidate structures that satisfy the theory are (the models of) the
well-formed signs. But if you think in terms of derivations, you'll still
perceive the arboreal backbone of a feature structure modlling (say)
a finite sentence as a derivation. (When I say arboreal backbone, I
mean the substructure obtained bt starting at the root and keeping all
the nodes that are reachable via edges labelled DAUGHTERS, HEAD-DAUGHTER,
etc.)

Whereas if you start at the root node and then THROW AWAY the
arboreal backbone (except the root), then what you are left with
is essentially what it means (CONTENT), what it sounds like (PHON),
and synatctic category it belongs to (CAT), which one might
argue is all you really care about.

You said:

 When I first started reading HPSG (in particular SWB), it felt to me
 like the tree was really a proof that the AVM of the root node could
 be composed out of the AVMs of the lexical entries in a manner
 consistant with the various constraints on the relationships between
 the lexical entries and the AVMs of the constituents that contain
 them.

 WHen I apply my own derivational biases to the system, the proof-like
 quality of the tree was almost a derivation (in fact can be read as
 such if you construe local constraints as "rules" and build from the
 bottom of the tree up. From my own biased point of view, then, the
 trees begin to look like old style derivational histories rather than
 structural representations.

and that is a completely natural way to view them. You can think of the
PS94 architecture as reifying derivations and then giving well-formedness
principles that separate good (reified) deriuations from bad.

I still think this is a good architecture, but that PS94 reified derivations
the wrong way.

>
I would have guessed, if asked, that the way HPSG was
going was to a system without trees at all, where there would only be
constraints from AVMs to AVMs, because of this derivational quality of
trees.
>>

If I understand you, Andreas Kathol's conception of grammar
architecture has had that property all along (since early-to-mid
1990's).

Carl