SLASH propagation vs. internal merge: Rewind

[Carl Pollard]

Hi Tibor,

[Sorry to clutter mailboxes but somehow I sent it the first time with
the last paragraph missing./

I haven't grasped the overall structure of the argument you are making
here. Is your position simply that SLASH values should be of type
_synsem_ (or maybe even _sign_), not _local_, and that position simply
IS the CTM couched in HPSG terms? If so, then are you advocating typer
_synsem_, or type _sign_, or just not doing HPSG?

>
I think that the Chamorro-case should be taken in the same way: The
phenomenon is not pervasive and in fact, we don't know whether languages
exist which actually mark an XP[SLASH Y[QUE ZP]].
>>

I take you you advocate allowing for the possibility that such languages do
exist?

>
These cases show that the 'standard theory of binding' in HPSG (PS
1994)   has problems with reconstruction and nothing more, as has
already been hinted  at by Stefan, who offers the following example:

> (2) a. Kennt er_*i Karls_i Freund?
>     b. [Karls_i Freund]_j kennt er_*i  _j.

While this example could be discussed away by assuming that Principle C d
oes
not exist,
>>

Right

>
we get the same problems with Principle A, witness

[1] Which picture of himself_i/j does John_i say Peter_j likes.

At first, [1] seems to show that Principle A in HPSG does it right,
because both indexations are surprisingly possible. Note, however,
that structurally identical examples in German and other languages
(one could call it *the majority*) do not allow the _i-coindexing. So
an appropriate cousin of [1  ] runs into the same problems as Stefan's
(2b).
>>

Does this remain true if LIKES is replaced by STOLE or DEFACED?

>
I have never worked on pseudo-cleft, but in order to make the contrast wo
rk,
one would have to assume that pseudo-clefts are derivationally related to
the ungrammatical examples provided by Bob.

> What he may do ___ is go home.
> *He may do go home.
>>


Consider also:

  What he may do next is he may just decide to forget the whole thing.

  What happpened next is (that) it started to hail.

(Thanks to Elizabeth Smith for pointing out that the postcopular
phrase in a pseudocleft can be a clause, even if the precopular
material either has no gap at all, or else has a gap that a clause
wouldn't fit into.)


>
The parasitic gaps seem to be the most interesting candidates. However,
Carl's recent email made it clear to me that they seem to be as problemat
ic
to HPSG as they possibly are to MP. Carl points out the ungrammaticality
of
[2] can be derived from the observation that "there is no way for the gap
 in
the upper filler (i.e. _i, TK) to get linked with the object gap."

[2] *[Without even reading _i]_j, [I don't know [[how many reports]_i [Ki
m
[[filed _i] _j]]]]

But how can this example be excluded in HPSG? There is no general --
possibly non-local -- condition to the effect that unlinked
filler-gap-relations are illicit in HPSG (there could be a coindexation
_i/_k, where all the information about _i is identical to all the
information about _k).
>>

This is true, but this is not so much a bug in a certain account of
UDC's as it is is a general HPSG architectural bug, one that was first
pointed out long ago by Tilman Hoehle: in general, unless there is an
explicit path-nonidentity constraint, there is always the possibility
that, in a feature structure where two distinct nodes root
structurally isomorphic substructures, one option is for those substructures
to be token-identical. In the present example, this results in an unwanted
identity between the gap in the filler and the object gap.

For example,

   A red car ran into a green car

has two structurally distinct analyses, one witht two distinct tokens
of CAR and one with only one. This clearly does nopt correspond to
a linguistic disinction we wish to make. But there does not seem to
be a theory that tells us which instances of structural distinctness
are the ones we are using to capture linguistic differences and which
are just unwanted artifacts of the formalism.

Carl