Inverse of :consist-of --- :consist or :consist-of-of?

[goodmami]

Some relations use "-of" in their apparently default form:

• :consist-of
• :on-behalf-of
• :out-of

If we invert these relations, do we remove the -of or add an additional -of? E.g., in "Gold makes up the ring"

(g / gold
   :consist (r / ring))

or

(g / gold
   :consist-of-of (r / ring))

In the guidelines underneath the relation inventory, it says "All relations above have inverses of the form :X-of", so my guess is the latter. Is this correct?

[nschneid]

I think :consist-of is an inverse that just happens to be more useful for annotators than :consist because of the syntax of the verb "consist" in English.

:prep-on-behalf-of and :prep-out-of are prepositional placeholders for semantic relations, and so are not inverses.

[goodmami]

Thanks for explaining. Would it make sense then to name relations with multiple words using underscores instead of hyphens (e.g. :prep-on_behalf_of and :prep-out_of), so that processors of AMR could rely on *-of being an inverted relation? Otherwise they'll have to maintain lists of these special relations.

[nschneid]

I'll raise this with the AMR design team, but given that we want to discourage :prep-X roles anyway, a more likely outcome is replacing those two with proper semantic relations.

In the meantime, I guess there needs to be a special case that any role starting with :prep- is not an inverse.

[nschneid]

I did find one :prep-X-of relation in the data that IS an inverse:

The U.S. is one of the few industrialized nations that does n't have a higher standard of regulation for the smooth , needle - like fibers such as crocidolite that are classified as amphobiles , according to Brooke T. Mossman , a professor of pathlogy at the University of Vermont College of Medicine .

...
:ARG0-of (h / have-03 :polarity - 
     :ARG1 (s2 / standard 
                 :ARG1-of (h2 / high-02 
                          :degree (m2 / more)) 
                 :prep-with-of (r / regulate-01 
                           :ARG1 (f2 / fiber 
...

This annotation is problematic and should be fixed.

[nschneid]

Ulf: the official inverse of :consist-of is :consist-of-of, though this never occurs in the corpus.

Can we remove :prep-on-behalf-of and :prep-out-of?

@ulfulf will look into whether there is an official list of inverses.

[uhermjakob]

All roles listed at http://www.isi.edu/~ulf/amr/lib/roles.html are primary (i.e. non-inverse) roles, in particular:

• :consist-of
• :prep-on-behalf-of
• :prep-out-of
• :instead-of (obsolete, superseded by frame instead-of-91)

That means that their inverses are :consist-of-of, :prep-on-behalf-of-of and :prep-out-of-of respectively. Number of times that such x-of-of forms currently occur in the AMR release corpus: 0. Please also note that the inverse of :domain is :mod.

:prep-with-of is indeed a valid inverse of :prep-with.

Ideally, all :prep-x and :conj-as-if will eventually be replaced by something more semantic. Compared to the early days of AMR, we have already made great progress to reduce the number of such :prep-x roles.

Concepts and roles use only dashes, never underscores. Allowing a mix of dashes and underscore will might make annotations more difficult for annotators and consistency checking.

AMR Dictionary entry for :consist-of updated to include a note on its inverse: http://www.isi.edu/~ulf/amr/lib/amr-dict.html#:consist-of

[nschneid]

Thanks @ulfulf. I've opened an issue in the AMR editor about sentence nw.wsj_0003.20, whose use of :prep-with-of looks wrong to me.

[nschneid]

The :prep-with-of is wrapped up in the treatment of "standard": #197

[goodmami]

Sorry to comment on a closed issue, but note the following:

http://www.isi.edu/~ulf/amr/lib/roles.html states the following under "Notes":

• For each role :role, there is an inverse role :role-of. The inverse of roles such as :consist-of is :consist etc.

Where, as @ulfulf pointed out (and documented at http://www.isi.edu/~ulf/amr/lib/amr-dict.html#:consist-of), the inverse is :consist-of-of.

[uhermjakob]

Thanks, Michael, for catching this inconsistency! I corrected http://www.isi.edu/~ulf/amr/lib/roles.html