Question formal status of trees in H&GPSG (fwd)

[Carl Pollard]

Hi Andrew,

>
So, could someone give me the short and dirty version of what GPSG and or
HPSG think about the formal status of trees.
>>

GPSG and HPSG aren't quite the same in this respect because in HPSG
the sorts of relationships that have usually been put into phrase
markers are included in the feature structures (which are the
theoretical embodiments of linguistic expressions conceived of as
structured objects). So for HPSG at least you have to ask whether
FEATURE STRUCTURES are manipulable or not. If you try to ask that,
then you have to get straight what it would mean to manipulate a
feature structure. If what you mean is: is everything in the feature
structure accessible from outside (in the sense of being referrable to
by some constraint), then TECHNICALLY the answer is yes, if you
formalize your grammars in the most carefully worked out logic that
has been developed specifically for formalizing HPSG grammars (namely
RSRL), since that logic is extremely expressive. I say TECHNICALLY
because, first, most people who write HPSG grammars aren't thinking
about how to formalize the constraints in logic, and second, because
even those who do think about that have pretheoretical notions of
locality that are reflected in the constraints that they actually
write.

For example, it would not be at all straightforward to write a lexical
entry for a verb that subcategorizes for a finite sentential
complement headed by a transitive verb whose direct object is
non3rdsng. Pretheoretically: you can't select the internal
mother-daughter structure of your complements. Theoretically, the
usual way to get this effect is to say that the datatype used for
selection of complements (and other valents) is one that doesn't have
that kind of information in it.  But if you discovered an Amazonian
language in which it appeared to be possible to do this, you could
still express it, say by adding a head feature OBJAGR to verbs whose
value was the empty list for intranstive verbs and the AGR of the
object for transitive ones.

So yes, HPSG feature structures are manipulable, as opposed to being
derivational histories.

Carl