[developers] Dropped arguments in DMRS

[Ann Copestake]

I'd like to come back to this point that Mike raised, since it also 
relates to ICONS and anaphora.  I decided not to include unexpressed 
arguments in DMRS from the ERG because:
a) clutter
b) shouldn't be needed in any application I could think of (they really 
shouldn't be needed in input to realization
and they can always be reintroduced via the SEM-I)
c) lack of clarity about precisely what they convey
Apologies if this has changed - also I know this issue has been 
discussed at some Summit meetings, and I haven't gone back over all such 
discussion, so sorry if the discussion is all old.

to expand on c) - there are (at least) the following cases (where I 
don't want to claim the division is clearcut):

i) syntactically optional argument unexpressed but (in the real world) 
something has to be there:
   either - with default understood argument - e.g., I already ate - see 
Fillmore (and also Lascarides + Copestake)
   or no default - e.g., I understand
ii) syntactically and denotationally optional - e.g. the donkey kicked - 
may or may not have kicked something,
- e.g., I ran (maybe a race)

There are also cases where there's an obligatory entity in the real 
world but it can't be expressed by a syntactic object - e.g., I dined
where I have to have eaten something but I'm not telling you what it is 
(I know one can say `I dined on salmon' or whatever, but according to 
the ERG, that's an adjunct not an argument).

[ There have also, in the past, also been cases where the ERG has a 
single predicate for two senses even though the verb is obligatorily 
(in)transitive in one.  I can't think of any examples offhand and this 
may have changed.  But that was a reason not to include them in the DMRS 
originally. ]

Current question - does this interact with pronouns?  If pronouns were 
possible with syntactically unexpressed arguments but not with the cases 
where the entity is denotationally required (like `dine'), then this 
would be a reason to keep them.  My intuition is that this is not the 
case for English.

I already ate.
It was spicy.

I already dined.
It was spicy.

For me, both are a bit iffy, though possible with the right context 
(I'll let you make up your own contexts ...)
`It was nice.' is OK (after both) but that could refer to the activity 
of dining/eating.  I think:

My stomach is full.
It was spicy.

is worse, but that may be because of the possible antecedent `stomach' 
giving a garden path or something.

Does anyone have cases where there seems to be a clearer difference?  Or 
another reason to include dropped arguments?

For Mike's examples, where the dropped arguments are necessarily 
coindexed, the case for representing this in the DMRS is more 
compelling.  So my thoughts are that the DMRS simply includes a node 
with a standard predicate (e.g., `unexp', but any name you all want as 
long as it doesn't clash with something already in use) which is a 
grammar pred and is linked to in the usual way.  So more-or-less 
equivalent to the zero pronoun analysis, assuming the zero pronouns 
don't have quantifiers.  Is this done with EDS?  I'll match terminology 
etc if so.

I would be interested in knowing whether this is something where 
subsequent pronominal reference to that is possible/easy or not.

If I do this, I can also implement the option to introduce nodes when 
there's a non-coindexed unexpressed argument.  I'll probably put the 
code in anyway but may not document the option unless someone comes up 
with a good argument!

All best,

Ann

On 05/01/2016 21:25, Michael Wayne Goodman wrote:
> Hi Ann,
>
> Thanks for sharing. I couldn't find the grammar at first because I was 
> looking in the LOGON tree instead of the separate LKB repository. If 
> others are searching, it's here: 
> http://svn.delph-in.net/lkb/trunk/src/data/dmrscomp/.
>
> I find DMRS more intuitive and more manageable than other *MRS 
> representations, so it's exciting to imagine a world where that is the 
> primary representation output by our grammars. I'm curious to see how 
> this works out with some larger grammars, but I can think of a couple 
> of challenges (based on my discussion in Singapore: 
> http://moin.delph-in.net/SingaporeMrsWellformedness).
>
> 1. We don't yet have a way to represent ICONS in DMRS
>
> 2. DMRS currently can't express coindexed dropped arguments (where in 
> MRS the 'i' variable of two arguments is the same; perhaps this can be 
> represented using ICONS instead, or by (re)introducing zero-pronouns)
>
> These are both difficulties with the resulting representation. I'm not 
> sure if there are other issues when implemented in the grammar. 
> Sometime soon it would be good to iron out these representational 
> wrinkles. Considering ICONS, I don't think we can just put a 
> post-post-slash label on a link (e.g. ARG1/NEQ/topic) because I don't 
> think ICONS follow normal dependency relations (Sanghoun could confirm).
>
>
> On Fri, Dec 18, 2015 at 9:46 AM Ann Copestake <aac10 at cl.cam.ac.uk 
> <mailto:aac10 at cl.cam.ac.uk>> wrote:
>
>     I have just checked in to the LKB svn repo a small grammar -
>     dmrscomp -
>     and some code that extracts simple DMRSs directly from the feature
>     structures produced by that grammar rather than going via MRS and
>     RMRS.
>     This is based on the mrscomp grammar (though with some clean up and
>     minor extension) - there's a fairly detailed README file. There are a
>     fair number of items on the TO-DO list - possibly the most
>     time-consuming one would be to make the generator code work with this
>     grammar, not because there's any big problem (that I can think of) but
>     because the generator is quite complicated.  There is also a
>     promise of
>     more detailed notes, which I will supply relatively soon, I hope -
>     this
>     was an interesting exercise in thinking through semantic composition.
>
>     If someone would like to collaborate on trying a similar exercise
>     with a
>     larger grammar, I'd be very interested.  It would help if it were a
>     grammar which already had the characteristic variable property, in
>     which
>     case I think the main part of the conversion should be fairly easy.
>
>     There are a number of potential advantages in constructing DMRS
>     directly, including the ability to construct a DMRS forest
>     directly from
>     a parse forest.  I would argue that it also enforces some notions of
>     semantic well-formedness more directly than is possible with MRS -
>     obviously including the (equivalent of) characteristic variable
>     property.  The semantic `fingerprint' of constructions can be
>     expressed
>     more simply, because DMRS removes much of the redundancy of MRS.  But,
>     of course, this is only interesting if we really can express
>     everything
>     we want to with DMRS.
>
>     All best,
>
>     Ann
>
>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20160110/ade056be/attachment.html>