[developers] smallish native DMRS grammar

Stephan Oepen oe at ifi.uio.no
Thu Jan 7 21:53:39 CET 2016

hi again, ann and joshua,

> But, of course, this is only interesting if we really can express everything
> we want to with DMRS.

yes, that sounds like the million-dollar question to me, and i believe
it connects to the relations to ‘semantic theory’ that joshua would
like us to be more explicit about.  but how do we go about seeking
answers, i.e. how to evaluate competing candidate analyses?

i recall when i first sketched the conversion to EDS to ann (sometime
2002, at the dining table of a soma loft) you were skeptical of the
distinguished variable property, because you wanted MRS to remain
compatible with a conventional analysis of intersective adjectives as
dog(x) ∧ fierce(x).  it would seem you have changed your point of view
on this design choice?

i understand we motivate adding an inherent event(uality) to adjective
and preposition semantics because of their potential for predicative
use in english; but i fail to see the parallel argument for the events
we postulate in adverb semantics (including operators like ‘not’ and
‘probably’) and non-surface critters like ‘poss’, ‘compound’, ‘comp’,
‘plus’, et al.  when pushed for explanations, i usually allude to the
requirements of composition in degree modification, e.g. ‘very
quickly’ and ‘really not’.  but this (arguably practical) argument
does not extend to the latter abstract predicates.

joshua (or anyone), what kinds of analyses (if any) does semantic
theory offer for degree modification of adjectives, adverbs, and such
(this is not a rhetorical question)?

as an aside: for the purpose of converting from MRS to dependency
graphs, i see no need to require a (unique) distinguished variable on
/all/ EPs; only logical (i.e. non-label) variables that occur in
argument positions need to be ‘introduced’ as the distinguished
variable of some EP.  in NorGram MRSs, for example, ‘poss’,
‘compound’, and such did not provide an ARG0 (possibly because we
failed to give helge an explanation of its utility :-), and that
worked just as well for transfer and EDS purposes.  in fact, if one
were to extend the disambiguating heuristics (currently for sets of
EPs that share one label) to cases like dog(x) ∧ fierce(x), i imagine
one could further relax the uniqueness requirement on distinguished

—so, in terms of expressivity in meaning representation languages,
DMRS seems more restrictive than ‘classic’ MRS, which can be an asset
or a disadvantage.  i suspect the ‘composite eventuality’ (two EPs
sharing the same ARG0) that joshua (maybe) would like in his
lushootseed grammar is technically parallel to an event-free analysis
of ‘fierce dog’: admissible in MRS, challenging in a variable-free
dependency graph, right?

best wishes, oe

More information about the developers mailing list