[erg] Intrinsic Arguments

Stephan Oepen oe at ifi.uio.no
Wed Feb 27 11:14:11 CET 2019


hi alexandre,

> The page is outdated? Is it a bug in ACE?

i would think neither of the above.  we might have to consider another
possibility here: the ERG does not (yet) fully implement uniqueness of
intrinsic arguments.  there are quite a few new derivational lexical
rules in 2018, and most of them instantiate the following type (from
‘lexrules.tdl’);

out_prefix_rule := basic_derivational_prefix_rule &
  [ INFLECTD -,
    DTR.SYNSEM [ LOCAL.CAT [ HEAD #head & verb & [ AUX - ],
                 VAL [ SUBJ #subj & [ FIRST synsem ],
                   SPR #spr,
                   SPEC #spec,
                   SPCMPS #spcmps,
                   COMPS *norplist* ],
                 MC #mc,
                 HC-LEX #hclex ] ],
    SYNSEM trans_subst & [ LOCAL [ CAT [ HEAD #head,
                     VAL [ SUBJ #subj,
                           SPR #spr,
                           SPEC #spec,
                           SPCMPS #spcmps,
                           COMPS < [ --SIND #cind,
                             OPT - ] > ],
                     MC #mc,
                     HC-LEX #hclex ],
                   CONT.HOOK.INDEX #index & non_conj_event ],
               LKEYS.KEYREL #keyrel & [ ARG0 #index,
                            ARG2 #cind ] ],
    C-CONT.RELS <! #keyrel !> ].

here, the C-CONT will introduce a new predication (the meaning
contribution of the verbal prefix), and it would seem the above type
unconditionally makes the prefix semantics share the intrinsic
argument (aka INDEX) of the verb.

i am no semanticist, but intuitively this analysis seems defensible to
me; the two predications could be interpreted as jointly constraining
the same eventuality .... somewhat parallel to a conventional analysis
like fierce'(x) ∧ dog'(x) for attributive adjectives.

uniqueness of intrinsic arguments is not a wellformedness requirement
in MRS per se.  it is rather an additional assumption that has
gradually gained support in DELPH-IN with growing interest in
conversion from MRS to semantic dependency graphs.  there is some
discussion of how to motivate this move for adjectives and adverbs on
ErgSemantics/Design; we should probably give some thought to how to
extend this argument to decomposed verb semantics.

technically, as the ERG semantic interface is in MRS, there is no
standard mechanism to enforce unique intrinsic arguments.  the
assertion you note on the ErgSemantics/Basic page, thus, is more of a
declaration of intention than a technical guarantee (much like many of
the other descriptions of expected outputs in the ErgSemantics
documentation).

thus, i would think consumers of ERG outputs will have to be somewhat
robust to the above configuration (and to other deviations from the
intended structures, of course).  i believe the conversion to EDS
should handle the configuration you notice gracefully and invoke the
disambiguation heuristics discussed on the EdsConversion page.  with a
bit of luck, that might even result in a decent EDS and DM, or?

best wishes, oe



More information about the erg mailing list