Attachment points

[Holger Schauer]

Hi,

I'm interested to hear comments on the question whether all units
identified in a text may possibly serve as attachment points or not.
Let me start off with an example, in which I take the size of
elementary discourse units to be clauses:

1a) Lexmark has announced its new LXZ printer,
1b) which has two seperate tanks for black/white and colour printing.
2)  The printer is capable of printing with a resolution of 600dpi.

... and so on.

I see two plausible attachment possibilities:

 - attach 2) to 1 (grouping both 1a and 1b)
 - attach 2) to 1a only.

Both times the relation is Elaboration-Object-Attribute.

I prefer to go with the second possibility, because the object (the
printer) elaborated on is contained in 1a. And although the object
would be "inferentially accessible" in 1, I strongly believe that 1b)
is clearly *not* elaborated on by 2.

However, if one does so, the question arises whether it is equally
allowed only to elaborate on the 1b). I think that this is possible,
yes, but I have some groans about that (see below):

Consider

2') The ink in the tanks will usually last for hundreds of print-outs.

This 2') now clearly only elaborates on the "tanks" from 1b), so I
have to conclude that this time we get [1a obj-attr [1b obj-attr 2]].

Now, let me talk about what makes my stomach hurt: first of all, this
second analysis crosses a sentential border, i.e., the analysis stands
in clear contradiction to the idea that the syntactic structure of 1
plays any role (yes, I know about RSTs focus on a "pre-realizational"
structure, however, this gives me far worse headaches). However, I
still think that it is justifiable, because the proposed analysis is
exactly the one that one would probably arrived at if 1 would go like
this:

1a') Lexmark presented their new LXZ printer.
1b') It has separate tanks for ...
2')  ...


However, note that one attribute of this little example is that the
further attribute-giving in 2) is easily covered by the obj-attr
relation between 1a and 1b. But what if the relation between 1a and 1b
would have been of a different kind, say purpose?

3a) Lexmark presented their new LXZ printer,
3b) to regain market-share in the low-price segment.
4)  The low-price segment has recently been dominated by HP.

Now, if we (naively) again use the above idea and annotate as [3a
purpose [3b elab-additional 4]], then we essentially say that the
information in 4) is also covered by the purpose-relation. And here is
the question: is that ``correct'', i.e. does it capture intuitions?
I think that this kind of structure becomes more implausible the more
information/units beneath the "purpose", such as
[3a purpose [3b elab [4-10]]]. But based on which criterion should one
decide when it makes more sense to say that [4-10] elaborates on 3
instead of on 3b and why when don't we say that for 4 alone?

For me, the question breaks down to whether constraints from the text
should be obeyed during annotations. That is, whether we should
restrict ourself to certain attachment points, like e.g. attach only
to "some" possible units like entire sentences or main clauses.

Any further comments?

TIA,

 Holger

--
Holger Schauer                         CLIF - Computational Linguistic Lab
                                       Freiburg University, Germany