Reentrancy in feature structures

[Luis Casillas]

On Wed, Jul 03, 2002 at 01:59:58PM -0400, Mike Maxwell wrote:

> As a lurker on this list, let me ask a couple naive questions.  First,
> it seems to me that the meaning of "identical" in Luis Casillas'
> definition is a crucial issue here.  Dredging something out of my old
> lisping days, I recall that Lisp made a distinction between (at least)
> two notions of "identical", often labeled "eq" and "equal".  One (I
> can't recall which) required identity of content (so e.g. two strings
> were equal under this definition if they were "spelled" the same), while
> the other required identity of reference (i.e. they were the same memory
> location).  I presume that condition 4 requires the latter (stronger)
> definition, correct?  In which case the equivalency really is just the
> same as re-entrancy, correct?

What I was suggesting with condition 4 was more like the "two different
strings that happen to be spelled the same", only that since feature
logic doesn't allow you to ever change them, you are guaranteed that
they will always be the same.  Feature logic is purely declarative, and
the nodes in feature structures are not memory cells which you can
change, creating aliasing problems in the case of reentrancy (of which
every Lisp book warns you).

> Second (and assuming the answer to the first question is "yes"), I can
> think of one obvious case where this sort of thing makes a difference,
> namely with co-reference.  E.g. in some language where verbs are marked
> for agreement with subject and object, you would get a reflexive form
> only where the subject and object agreement features were identical in
> the strong sense.  (Otherwise you couldn't distinguish the translation
> equivalent of "He saw him" from "He saw himself.")

Yes, this is one case I was thinking about, coindexation.  But can't
you do at least the morphology of the reflexives simply by having them
require type identity between the subject and object indices, while
leaving the regular transitive structures unconstrained?

The semantics is another issue, since the feature structures are getting
mapped to semantic objects in a way that is sensitive to reentrancy;
the choice between token and type identity for indices is, in a sense,
controlling the instantiation of variables.  I.e. if the indices for
subject and object are token identical, we instantiate `see(i,i)', while
if they are distinct we get `see(i,j)' even if both i and j are 3sg.M.

Which probably means that under my proposal, you'd want to add a LETTER
feature to indices, taking atomic values like i, j, k, and add a
constraint to non-reflexive constructions requiring different values
of LETTER for their arguments. Then in the semantics, we map index
structures with the same LETTER to the same variable, and those with
different LETTER to different values.  Yes, a hack.

Coincidentally, I think the only time the distinction between type and
token identity came up when I took LFG was with reflexives.

--
Luis Casillas
Department of Linguistics
Stanford University
http://www.stanford.edu/~casillas/