two interpretations of direct object of 'risk'

[nschneid]

In an EMNLP talk, Gerald Penn points out that the direct object of the verb risk can mean either of two things: an asset (good) that is put in jeopardy, or a negative consequence (bad) that becomes a danger. He found that conflating these two readings up is problematic for summarization.

It appears that neither PropBank nor VerbNet distinguish these two readings: there is a single role for the direct object, ARG1 or Theme. Perhaps this is because the two versions of this argument never appear together in a sentence.

Is this a problem for AMR? Here are some examples returned by a release search:

Asset reading

good savings accounts where capital is not at risk

(a2 / account
    :mod (s2 / save-01)
    :mod (g2 / good)
    :location-of (r / risk-01 :polarity -
                   :ARG1 (c / capital)))

risking public health

(r / risk-01
    :ARG0 g
    :ARG1 (h / health
               :mod (p / public)))

Dangerous consequence reading

risk surveillance or arrest

(r2 / risk-01
         :ARG0 p
         :ARG1 (o3 / or
                 :op1 (s / surveil-01
                        :ARG1 p)
                 :op2 (a3 / arrest-01
                        :ARG1 p)))

the serious earthquake risk in the Ngawa region

(r / risk-01
   :ARG1 (e / earthquake)
   :mod (s4 / serious)
   :location (l2 / local-region :wiki "Ngawa_Tibetan_and_Qiang_Autonomous_Prefecture"
           :name (n2 / name :op1 "Ngawa")))

Using :mod, not risk-01

relatively low risk of serious infections

(r2 / risk
      :mod (l / low
              :degree (r3 / relative))
      :mod (t / thing
              :ARG2-of (i / infect-01)
              :mod (s / serious)))

health risks are minimal

 (m / minimal
     :domain (r / risk
               :mod (h / health)))

[nschneid]

In Unified PropBank, I think the two roles ought to be distinguished anyway for risk of vs. risk to/at risk. Cf. danger of/to and in danger; peril of/to and in peril; jeopardy of/to and in jeopardy.

[timjogorman]

Here's a link to the Fillmore paper: http://www.icsi.berkeley.edu/pubs/ai/towarda92.pdf

[nschneid]

Decision: This probably needs separate PropBank roles.