search

amrisi / amr-guidelines

244 stars 87 forks source link

Re-ification of :mod (?) #117

Open uhermjakob opened 10 years ago

uhermjakob commented 10 years ago

There were a few tricky sentences in the SDL section of sentences that annotators (including me) struggled with and that kept getting AMR Checker error messages. I finally resorted to re-ifying :mod. 

Jewish people and half-Jewish people
(a / and
  :op1 (p / person
         :mod (r / religious-group :name (n / name :op1 "Jew")))
  :op2 (p2 / person
         :ARG1-of (h / have-mod-91
                    :ARG2 r
                    :degree (h2 / half))))

Similarly: Brits and non-Brits, with (have-mod-91 :polarity -)

AMR search command: rs have-mod-91 (will yield 4 instances)

Essentially the same, but maybe more dignified might be have-domain-91 with swapped args.

Any better ideas?

nschneid commented 10 years ago

Under this proposal, I suppose "semi-automatic weapon" would be

(w / weapon
   :mod (a / automatic
           :degree (s / semi)))

whereas "semi-democratic election" would be

(e / elect-01
   :ARG1-of (h / have-mod-91
               :ARG2 (d / democracy)
               :degree (s / semi)))

right? I can't say I'm wild about how different those look, but this may be the best option nevertheless.

timjogorman commented 10 years ago

I'd love to have this, but I'm a little unclear about how much we should use it -- especially with negation. To modify Nathan's example above, could "undemocratic election" be the following?:

(e / elect-01
      :ARG1-of (h / have-mod-91 :polarity -
            :ARG2 (d / democracy)))

Many negated adjectives seem to exist on some scale between meaning "X has the attribute of not-Y" and meaning "X does not have the attribute Y", with lexicalized negations like "my stomach is uneasy" at one end and equational negations like "john is not a lawyer" at another. I like that we've really lumped those senses together so far, by not allowing "be". Have-mod-91 would essentially add what amounts of an attributive copula, and so I'd suspect that the some annotators would thus be tempted to annotate "John is not a laywer" as:

(l / lawyer
      :ARG2-of (h2 / have-mod-91 :polarity -
            :ARG1 (p2 / person :name (n / name :op1 "John"))))

Do we want to allow this? If not, can we explicitly discourage it?