[developers] smallish native DMRS grammar

Stephan Oepen oe at ifi.uio.no
Wed Jan 6 23:00:03 CET 2016


> (Sorry, I know oe uses different terminology, but I can't remember what it
> is at the moment.)

i first developed the conversion from MRS to variable-free semantic
dependency graphs, called Elementary Dependency Structures (EDS) in
2002; it appears i first coined the term ‘distinguished variable’ in
2006.  i would recommend you look at the EdsTop page on the wiki and
Oepen et al., (2002; TLT) and Oepen & Lønning (2006; LREC) as
additional background.

i would be interested in understanding better the relationship between
EDS and DMRS.  in Copestake (2009; EACL), you say that DMRS extends
EDS, but you do not discuss specific differences.  as far as i can
tell, the basic construction is the same: one dependency node per EP,
dependency edges for each argument, picking target nodes according to
their distinguished variable or label, with a set of disambiguating

DMRS overlays the EDS with information about the label topology, i.e.
it adds the ‘post’ annotations on dependency edges as well as some
undirected edges with only ‘post’ labels, for some constructions.
these additions overcome the lossy nature of conversion to EDS, i.e.
enable conversion back to the original MRS.  they also add some
complexity that may be undesirable for some applications: for
instance, attributive vs. predicative usages of adjectives differ in
their ‘post’ annotation, e.g. ARG1/eq vs. ARG1/neq for ‘the fierce
dog’ vs. ‘the dog is fierce’.  which variant is better suited for
specific tasks i find an intriguing open question.

on a more technical level, EDS provides some additional facilities
(some historic), e.g. something akin to predicate modification
(‘nearly every’), more elaborate disambiguation heuristics (possibly
relevant only with deviant MRSs), a choice of including quantifiers or
not, elimination of messages, output to various triple formats,
computation of dependency precision and recall, and such.

seeing there is a lot of common ground, why introduce parallel
terminology?  and would you agree that it is fair to encourage people
to cite both Oepen & Lønning (2006) and Copestake (2009) as relevant
references for DMRS?  truth be told, i would welcome additional
visibility for my original work on relating MRSs and dependency

best wishes, oe

More information about the developers mailing list