[developers] Reg:- Possibility of conversion from EDS format to MRS

[Stephan Oepen]

dear prateek (if i may),

i am intrigued to hear about your use of the LOGON infrastructure, EDS
meaning represantations, and ERG generation for machine translation.  i
have been part of the development of all three resources and will be happy
to try and assist—even more so since i believe you are asking a question i
have wanted to investigate myself earlier.

EDS (Oepen & Lønning, 2006; LREC) is designed as a /lossy/ reduction of
full MRSs, leaving out most of the underspecified scope information that,
arguably, is among the more intricate parts of ERG semantics and that,
arguably, makes a comparatively smaller contribution to ‘core meaning’.
 thus, among the original EDS goals was simplification (removing
MRS-specific ‘technicalities’), in order to relate to dependency-based
approaches to meaning representation and to enable practical downstream
tasks.

to give a few examples, full ERG MRSs require that propositional arguments
are scopally sub-ordinated by a =q handle constraint; quantifiers always
bind their restrictor via =q, leave their body unbound, and never share
their label; intersective modifiers in attributive use (‘the fierce dog’)
share the label of their head, but in predicative use (‘the dog is fierce’)
must not do so.

for an input MRS to succeed in generation, the above wellformedness
constraints (and others) must be satisfied, and the generator therefore is
not very robust to ‘deviant’ input structures; underspecification of the
label topology is often not possible.  working towards generation from
(formally less complex) EDSs, i expect, might lower the burden for creators
of input semantics for ERG generation.

there is no pre-existing converter from EDS to MRS to date, and for the
above reasons such conversion will require some additional (scope-related)
information, beyond what is explicit in the EDS.  however, i expect that
the bulk of the missing information could be derived (or inferred, if you
will) automatically from the ERG lexicon and SEM-I, for example scopal
argument positions and classes of intersective modifiers.  it could be very
interesting to try and validate that hypothesis :-).

i can imagine at least two technical solutions one could pursue here.
 personally, i might envision a two-stage approach: first, trivially
rewrite an EDS into an incomplete ‘pseudo-’MRS, with one EP per node,
logical variables introduced according to the ‘distinguished variable’
notion of Oepen & Lønning (2006), and one role per dependency edge.
 second, one could devise MRS transfer rules to introduce scope-related
information, e.g. labels, scopal arguments, and handle constraints as in my
examples above.

i imagine many such rules could be auto-generated from mining the ERG
lexicon and SEM-I, and both the LKB and ACE generators already allow the
application of a pre-generation ‘fix-up’ transfer grammar.  mid- to
long-term, these transfer rules could become an integral (if maybe
optional) part of the ERG generation set-up.

an alternate (if conceptually probably equivalent) solution would be to
augment the EDS graph into what Copestake (2009; EACL) calls Dependency
MRS, or DMRS.  DMRS is an extension to EDS, in that dependency edges are
overlayed with additional scope-related information, e.g. label equality or
inequality for the ARG1 of an attributively or predicatively used
adjective, respectively (my ‘fierce’ example above).  in addition to these
overlay annotations on edges also present in the EDS, DMRS occasionally
introduces additional undirected eges, to record label sharing between EPs
(aka nodes in the graph) where there is no argument relation to which an
overlay could be attached.

unlike EDS, the DMRS design capitalizes on lossless bi-directional
conversion, and the LKB at least includes code to create a standard MRS
from a DMRS; it would presumably be a small modification to the generator
to call out to the conversion when presented with an input DMRS.  in this
scenario, however, the injection of scope-related information missing from
your original EDSs could not be easily implemented semi-declaratively as
transfer rules, as it would have to happen while going from EDS to DMRS,
i.e. prior to conversion to MRS (where the LOGON transfer machinery works).

i hope the above makes sense to you, and i would be curious to learn more
about your project at large and about which path into ERG generation you
decide to pursue.

best wishes, oe


On Thursday, November 5, 2015, Emily M. Bender <ebender at uw.edu> wrote:

> Dear Prateek,
>
> Can you tell me some more about how you are using the LOGON
> system?  MRS format is the native format of the ERG; if you have EDS,
> representations, they have probably been converted from MRS.
>
> Or are you producing EDS directly for some language?
>
> Emily
>
>
> On Thu, Nov 5, 2015 at 3:55 AM, Prateek Saxena <
> prateeksaxena2809 at gmail.com
> <javascript:_e(%7B%7D,'cvml','prateeksaxena2809 at gmail.com');>> wrote:
>
>> Hello ma'am ,
>>
>> We have been using the LOGON system in relation to parsing indian
>> languages, in course of which we require a tool for conversion of a parse
>> in EDS format to a parse in MRS format. Is there such a tool available for
>> automation of the aforementioned task.
>>
>> Thank you,
>> Prateek Saxena,
>> IIIT Hyderabad,
>> India
>>
>
>
>
> --
> Emily M. Bender
> Professor, Department of Linguistics
> Check out CLMS on facebook! http://www.facebook.com/uwclma
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20151106/38064c37/attachment.html>