Labels of RSTR at NP level in DMRS


#1

My understanding of MRS is that we make sure to hide the label of the N’ (what ends up qeq’d by RSTR on the quantifier relation) at the NP level, and that this is why the head-specifier rule is non-head compositional.

I’m curious what happens when composition is done directly in DMRS. The noun’s index is still exposed at the NP level (has to be), so could something make an /eq link to that index and effectively tuck itself inside the RSTR value of the quantifier? If not, what prevents this? If so, is it a feature or a bug?

On the feature side—there are the examples we found in Turkish where adjectives appear to be attaching outside the quantifier in the syntax. Also probably helpful for Wambaya!

On the bug side—there’s clearly a bunch of reasons we’ve been so careful to hide those handles in MRS all these years…


#2

I’m probably misremembering but wasn’t something like this what
the failed LSA abstract was about?


#3

Ah indeed! So, there we thought it was a feature. But then maybe it’s worth thinking about why this was disallowed in MRS?


#4

as far as I can remember, it’s not that we specifically
disallowed it in MRS, but that it’s a consequence of only exposing
INDEX, LTOP and XARG and wanting the NPs to be able to move around
in the underspecified scope tree. Given that quantifiers are
scopal, it would be rather odd to make the NP’s LTOP the LTOP of
the N’.

If we wanted the DMRS-type behaviour in MRS, we could (I think)
have made the LTOP of the quantifiers be the LTOP of the N’, even
though it’s weird, but we would have to make the combination of
the NP with another constituent a special case. In DMRS,
combination with an NP is anyway (sort of) a special case - the
neq link. NB - this is making explicit something we don’t make
explicit in MRS composition - i.e., the absence of equality. MRS
relies on a sort of pseudo-default behaviour, that all things
which are not explicitly equated are assumed “at the end” to be
non-equivalent (sometimes, loosely, referred to as Skolemization).

I don’t remember the details of the proposed approach for Turkish
with DMRS, but the combination of the adjective with the NP
doesn’t necessarily need to be a special case of a special case.
If the syntax allows the adjective plus NP combination (which
presumably it doesn’t for English), but the adjective is something
that says it has to have an /eq link, then it will get into the
restrictor.

I may have garbled some of this, but it’s one of those cases
where either I respond quickly or not at all …


#5

The idea with Turkish was that the adjective, though attaching to the NP (rather than N’) could use an /eq link to the NP and thus pick up the label associated with the INDEX (the noun’s), I think.

The only reason I can think of to ‘hide’ the LTOP of the noun (while exposing is INDEX) is that in the 2001 algebra, head-comps et al insist on LTOP identity between the head & non-head daughters. I guess this isn’t true in the DMRS algebra?


#6

I think saying that the LTOP identity of the daughters on rules
is the only reason to `hide’ the N’ LTOP is partly true but is
really getting things the wrong way round.

If one assumes that quantifiers are scopal (which they are, of
course), then the LTOP of the NP should not be the LTOP of the
N’. So the LTOP of the N’ is not going to be available. However,
because in MRS we don’t set the LTOP of the quantifier and don’t
use it, we could have exceptionally said that the LTOP of the NP
is the LTOP of the N’. Then we would also have to have an
exception to the principle that non-scopal combination (e.g., with
the head-complements rule) uniformly involves unification of
LTOPs.

As I suggested in the previous message, we could have done these
things (I think), but it seems messy. If you look at the partial
tree for the NP it’s weird, because one has a qeq from the RSTR to
an internal component which is the LTOP. It also prevents
modification of quantifiers by means of the LTOP (for `nearly
every’ etc).

In DMRS there are no labels (in the MRS sense) as such.
Obviously we can formalize via MRS, but we don’t have to, and
these days I wouldn’t. The /eq link is directly interpreted as
saying that any fully scoped trees must meet the constraint that
the components are at the same level and are conjoined. In
DMRS, there are three classes of combination operations, rather
than the two in MRS: qeq, eq and neq. qeq is much as in MRS
(though a bit easier to describe because we don’t have to talk
about holes and labels), eq is as above, and neq says `I don’t
know where you go in the scope tree, but you’re not eq and you’re
not qeq’. In a sense, this is a form of typing. qeq arguments go
for LTOP, eq and neq things go for INDEX.

In the sample grammar, things that may take an N’ or an NP have
have their arguments typed lexically as ueq, but there shouldn’t
be any cases where something ends up with ueq once the argument is
instantiated.

The thing I can’t remember about the Turkish analysis (and maybe we
didn’t work it out precisely) is under exactly what conditions one
ends up with an neq link. Obviously it won’t work if it’s every
time one combines something with an NP, which is what I assume for
the dmrscomp grammar (I think).