[rmrs] Questions on DMRS

[Michael Wayne Goodman]

Hi again,

Some of my previous questions were answered in Saarland this year, but I
have a few remaining regarding LTOP links:

(1) Why are both the pre- and post-slash content of LTOP links "NIL"?

(2) Related to (1), how do we tell the difference between MRSs where:
 - the LTOP label is equated with some EP's label
 - the LTOP label (or hole, rather) is QEQed to some EP's label
 - the LTOP label is neither a label nor a hole in the MRS*

* This last case might not be well-formed. It was briefly the case that
Matrix grammars behaved this way, but they may have just been missing the
appropriate QEQ.

(3) Why bother with multiple LTOP links? Shouldn't there just be one
pointing to the head of the set of nodes in a label-equality-set? The */EQ
links in that set should cover the rest of the LTOP associations.

Thanks in advance


On Fri, Mar 8, 2013 at 6:56 PM, Michael Wayne Goodman <
goodmami at u.washington.edu> wrote:

> Hi RMRS-list,
>
> I'm working on adding DMRS reading/writing support to my Python MRS
> library. I've read through the Slacker Semantics paper and comments in
> the Lisp code, but I have some remaining questions. Sorry this email
> turned out quite long; even partial responses would be appreciated.
>
> * What does the paper mean by "head", as in Sec. 4.2 "Head selection
> in the qeq graph". Is this like the heads of other dependency grammar
> formalisms? It seems to, at least, be different from syntactic heads.
>
> * According to the lisp code, if an LTOP labels of a set of EPs, then
> "the LTOP node in the DMRS is linked to each head predicate in that
> set". This "LTOP node" is implicit, yes? I see its ID is 0, but I
> don't see the node itself being output anywhere.
>
> * Regarding the LTOP link, should it be an undirected /EQ link? When I
> try with the LKB, I see a link with no pre- nor post-slash content (or
> rather, NIL):
>
> <link from='0' to='10004'><rargname>NIL</rargname><post>NIL</post></link>
>
> * Furthermore, why is there textual content "NIL" rather than an empty
> element <rargname/>? Would an empty element be interpreted the same?
> As it is, this seems like it would break if, ${deity}-forbid, some
> grammar developer defined a "NIL" role for EPs, such that "NIL" is
> actually the rargname. E.g.:
> [ "_example_n_1_rel" LBL: h3 ARG0:x4 NIL:h5 ]
>
> * Also related to LTOP links, when might we see the LTOP being linked
> to more than one node? Admittedly I'm not very good at coming up with
> examples, but I can't imagine such a sentence. Of course there's
> things like "the dog barked loudly", where "loudly" and "barked" share
> the label identified with LTOP, but in this case there would only be
> one link going to the head of the label-equality-set: "loudly".
>
> * I'm also interested in DMRS-MRS conversion, and so I'd like to
> recover the INDEX. I currently look for EPs that have no arguments
> within their label-eq-set, and are not the argument of an EP outside
> their label-eq-set. This by itself mostly works, except it can't
> select between "bark" and "maybe" in "some dogs bark maybe". I can
> exclude "maybe" if I look for EPs that don't have a hole argument
> QEQing some other EP's label, but then I fail to select "and" in "some
> dogs bark and sleep" (since L-HNDL and R-HNDL are hole arguments on
> "and"). I could then refine this so it ignores EPs QEQing some other
> EP only if the other EP's ARG0 isn't also an argument of the original
> EP (e.g. "bark"'s label is QEQed from "and"'s L-HNDL, but its ARG0 is
> also "and"'s L-INDEX), but now it feels like a hack. Is there a more
> reliable way to do this, preferably without loading the grammar and/or
> the SEM-I?
>
> Thank you,
>
> --
> -Michael Wayne Goodman
>



-- 
-Michael Wayne Goodman
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/rmrs/attachments/20130918/ca2d7d3b/attachment.html>