search

amrisi / amr-guidelines

246 stars 86 forks source link

Coreference and :domain #189

Open nschneid opened 8 years ago

nschneid commented 8 years ago

Johan Bos's SemEval poster pointed out a problem with the following gold AMR:

They are thugs and deserve a bullet.

(a / and
  :op1 (t / thug
      :domain (t2 / they))
  :op2 (d / deserve-01
      :ARG0 t
      :ARG1 (b / bullet)))

The problem is that it is "they" (t2) who deserve a bullet, not thugs in general. Although t is an instance of thug, so it's not clear if it's doing anything useful.

Where a pronoun is linked to a predicate nominal ("They are thugs"), it seems weird that we have a variable for the predicate nominal concept. We're playing fast and loose with the / operator: normally it means :instance-of, but it's actually doing something different with pronoun concepts (because the pronoun helps identify the referent; it is not a kind). Having both t and t2 seems superfluous.

Just brainstorming here—maybe we really want (t = they / thug), where = links a variable to its linguistic manifestation as a pronoun, when that link is asserted by the sentence?

Or is this the same thing as "Clark Kent is Superman"—and we just have to know that, whichever variable is repeated elsewhere, they are equivalent at some level?

uhermjakob commented 8 years ago

I think that in this specific case, based on context, the AMR should have been: They (black criminals) are thugs and they (black criminals) deserve a bullet.

(a / and
  :op1 (t / thug
      :domain (t2 / they))
  :op2 (d / deserve-01
      :ARG0 t2
      :ARG1 (b / bullet)))

In this AMR, they (black criminals) and thug (presumably including white thugs) don't refer to the same entity. I agree with Nathan that the / operator does not always mean :instance-of, particularly not for pronouns. Generally, constructions such as (h / he :domain ...) don't make sense for that reason.