ARG-ST as a head feature

[Martin Jansche]

Hi Carl,

On Mon, 29 Jan 2001, Carl Pollard wrote:

> This is the difference from Bayer: the coordination/neutralization
> lattice is IN ADDITION TO the conjunction/disjunction lattice.

That is a technical difference indeed, and I think I understand the
reason for it; more on this below.  But you still agree with Bayer
that there is a fundamental difference between valence features and
other features.  This will have empirical consequences that make it
quite different from 1994 HPSG.  But how does it differ from Bayer and
Johnson empirically?

> The two lattices have very different structures: the former is
> Alex(Pow(P)) (where the subsumption order on P is ignored) and the
> latter is Smyth(P) where the subsumption order on P is retained.

How does one decide what values go into set P?

> This differs from Bayer/Johnson because they CONFLATE
> conjunction/disjunction with neutralization/coordination,

I'm guessing that they would insist that this is intentional and
desirable -- not only is this not a problem in CG, it is the essence
of their analysis.

> with the undesirable consequences noted by Heylen. (Roughly, you
> can't describe a man with a dissociated personality, where one of
> the personalities is a vegetarian optometrist and the other is a
> coke-snorting gunrunner: the gunrunner gets turned into a
> vegetarian. In case the metaphor is too opaque, being a vegetarian
> optometrist is conjunction; being schizophrenic is
> neutralization.)

I'm assuming you're referring to the discussion of e.g. "her" (which
is ambiguous, not neutral) etc. at the end of ch. 9 of Heylen's
dissertation (also in a separate paper).  Very briefly, in CG it is
possible (and in this case mandated by empirical data) to assign two
different types to "her" (personal pron. vs. possessive pron.).  Not
so in the HPSG lexicon: basically, there is only one lexical entry for
each word, though it may be disjunctive.  This can lead to
overgeneration in conjunction constructions, as Heylen notes.

You're proposing a solution to this problem (note that the problem is
specific to a particular formulation of the HPSG framework).  That's
why I was saying that it seems that now you've caught up with CG
again.

(BTW, why not go for a simple solution like allowing multiple distinct
lexical entries?  Would we see an empirical difference?)

> The treatment of ARG-ST as a head feature using the freeze construct
> is not very much like anything I know of in categorial grammar.

CG does have mechanisms for locking and unlocking things, though they
have been used for different purposes.

> (But the pretheoretic analogy between valence features
> and categorial slashes has been made explicit since the inception of
> HPSG.)

The analogy had always been left at the intuitive level.  The
parallelism broke down as soon as you thought about contravariance of
what's "under" the slash and the terms ("semantics") associated with
CG slash-types, which correspond to abstraction.

> To call this inventing categorial grammar is a little bit like
> noting the equivalence between boolean rings and boolean algebras,
> and then concluding that ring theory is just a reinvention of
> lattice theory!

I was trying to epxress concern about how you would market your
proposal to the formal grammar community.  What your treatment is
useful for could be done in CG already.  My concern is not so much of
a theoretical kind (it has been demonstrated that there are
Turing-complete versions of CG, so you can state any HPSG theory in
them, and vice versa), but a practical: how easy is it to do things?
what price do you pay for it in terms of technical machinery?  It
seems that CG is still strong in both respects.

But in the end it is the empirical side that counts: can you do more
than CG can do?  (There are lots of other flaws with simplistic
accounts of coordination in CG that seem difficult to solve.)  Your
proposal so far allows a treatment of valence vs. non-valence features
in coordinate constructions that is very similar to CG, and you also
separate the mechanism for that from the one that allows you to define
the lexical component.  Is there something you can do easily now that
CG cannot?

Regards,

- martin