ARG-ST as a head feature

[J P Blevins]

>Not that I really want to adopt this
> >>
>
>Why not? The locality argument against doing this is not very
>persuasive, since the existing feature geometries already allow
>selection for things that are never selected for (e.g. parts of
>CONTENT such as QUANTS, having a certain MOD value, etc.).
>
>Regards,
>
>Carl

To follow up on Carl's point, it has always seemed to me that concerns
about locality stem from an essentially configurational view of locality,
which is in many ways alien to the spirit of approaches like HPSG.  The
basic idea appears to be that features are accessible if they are present
in a particular location and inaccessible otherwise.  However, there are of
course  other strategies for access control.  One simple model, with
obvious application to inheritance-based grammar formalisms, is the
private/protected/public distinction within object-oriented derivatives of
C.  A straightforward application to feature structures would classify
particular features, such as the CONTENT features mentioned by Carl, as
syntactically inaccessible (i.e., private). Another class, which would
include the traditional HEAD features, would be classified as accessible to
the head daughter (i.e., the analogue of protected).  A third class would
be accessible to all daughters and siblings of a node (i.e.,
public).   Within a simple system like this, a head's SUBJ and COMPS
valence features could be public accessor features that, in effect, permit
syntactic subjects and complements to modify, respectively, the FIRST
element and REST sublist of the head's ARG-S.  If the ARG-S feature itself
is classified as protected, it will only be directly accessible to a head
daughter, so that traditional cancellation will prevent any syntactic
dependent from accessing or modifying the ARG-S|REST value of a VP or S, or
the ARG-S|FIRST value of an S.

In this way, ARG-S could be shared without any relaxation of locality.  The
same sort of approach to access control would allow an HPSG grammar to run
in 'subsumption mode', in which ARG-S values would become progressively
more informative as you move 'upwards' from a lexical verb to the VP and S
that it heads.  It is fairly straightforward to modify a classical LFG
grammar to run in subsumption mode, but difficult or even impossible to do
this with a standard HPSG grammar.  (FWIW, I have a draft that outlines
some arguments for why one might want to do this, available at
http://www.rceal.cam.ac.uk/jblevins.html#fgb).

-Jim


___________________________________________
Research Centre for English and Applied Linguistics
University of Cambridge          jpb39 at cam.ac.uk
Cambridge, CB2 1QA, UK      www.rceal.cam.ac.uk