No subject

[Malouf R.]

Approve: 8789_18953
Received: from mailserver.cog.brown.edu (www.cog.brown.edu [128.148.208.165])
	by lists.Stanford.EDU (8.11.3/8.11.3) with ESMTP id fAHFbNM27716
	for <hpsg-l at lists.Stanford.EDU>; Sat, 17 Nov 2001 07:37:23 -0800 (PST)
Received: from localhost (mj1 at localhost)
	by mailserver.cog.brown.edu (8.9.3+Sun/8.9.1) with ESMTP id KAA12739;
	Sat, 17 Nov 2001 10:26:37 -0500 (EST)
Date: Sat, 17 Nov 2001 10:26:36 -0500 (EST)
From: Mark Johnson <mj1 at mailserver.cog.brown.edu>
To: "Stephen M. Wechsler" <wechsler at mail.utexas.edu>
cc: hpsg-l at lists.Stanford.EDU
Subject: Re: Default unification: online vs. offline
In-Reply-To: <p05010406b81b54f0de70@[146.6.51.17]>
Message-ID: <Pine.GSO.4.05.10111171011220.12707-100000 at mailserver.cog.brown.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII



On Fri, 16 Nov 2001, Stephen M. Wechsler wrote:

> This message from Mark Johnson, responding to my question about the
> feature ANY, may have bounced from the list:
>
> At 12:20 PM -0500 11/16/01, Mark Johnson wrote:
> >Hi Steve,
> >
> >I don't know whether HPSG still contains ANY features (either directly or
> >indirectly), but you're right that it causes a certain amount of
> >theoretical grief.  The easiest way of understanding ANY features involves
> >the notion of two different evaluation ``times'' (you solve the
> >constraints ignoring the ANYs, then check to see if the minimal model
> >solutions satisfy the constraints with the ANY features), which is
> >horribly non-declarative.
> >
> >I wrote a number of papers describing how to express ANY constraints in
> >various kinds of default logics.  To the extent to which we can give a
> >declarative semantics to these default logics, we have a declarative
> >semantics for ANY constraints.  Unfortunately, the semantics of
> >constraint logics is itself not exactly a picture of clarity:  As Samuel
> >Johnson (I think it was him) said about the dog walking on its hind legs,
> >it's amazing that it can be done at all, but it's still not what I'd call
> >a success.
> >
> >However, I think it's also worth asking what the implications are of not
> >having a nice simple declarative semantics.  While having a simple
> >declarative semantics is undoubtedly better than not having one, having a
> >simple declarative semantics has little to do with the insightfulness of
> >linguistic analyses you (can) use the theory to express, or the
> >computational complexity of parsing or generating using that theory.
> >
> >In other words, I guess I now believe that if having ANY values or
> >equivalent devices permits simpler, more insightful linguistic analyses,
> >then by all means use them, and let the mathematicans worry about the best
> >way to model them formally,
> >
> >Best
> >
> >Mark
>
> This is what I wanted to hear.
>
> Our of curiosity, then, why is this mechanism not considered a problem for LFG?
>
> Steve
>

I guess it depends on what you mean by "problem".    Worrying about what
ANY values might mean semantically kept me and a few others busy for a
while, and it wouldn't surprise me if there aren't better ways of
modelling ANY values semantically.

On the other hand, it's clear enough what ANY values mean that grammar
writers and implementors don't have any problem using them or implementing
them.  There are far greater sources of computational inefficiency in
standard implementations of both LFG and HPSG than ANY values.  In other
words, it's not clear that apart from aesthetics, there is
any reason to worry about the fact that we can't give ANY values a simple,
clean declarative semantics.

As one of my statistican colleagues pointed out to me, high precision does
not imply even modest accuracy.  A grammar formalism equipped with an
explicit formal semantics is very precise, but that has little to do with
whether it is linguistically accurate.  And linguistic accuracy (rather
than formal precision) should be our goal, I think.

Best,

Mark