Ambiguous reification of ":poss"

[flipz357]

Reification of ":poss" is different from all other reification rules since it seems there are two options, using either "own-01" or "have-03". I wonder what should be preferred when reifying "poss" relation, or can it make sense to introduce an extra amr concept, such as "possession-91" to clear the ambiguity?

[nschneid]

I don't recall why we say there are two reifications of :poss, but I'm guessing in practice :poss is annotated for attributive possession constructions and own-01 and have-03 are the most typical ways to indicate predicative possession, following the verb. I guess you'd have to choose one for "The car is his". Maybe own-01 is more specific?

[flipz357]

Yup, own-01 seems more specific. Maybe it's also reflected in PropBank, where roles for have tend to be more ambiguous:

have:

Arg0-pag: owner (VN Roles: 100.1-pivot, 39.4-agent) Arg1-ppt: possession (VN Roles: 100.1-theme, 39.4-patient)

vs own:

Arg0-PAG: owner (VN Roles: 100-pivot) Arg1-PPT: possession (VN Roles: 100-theme)

Don't know if that's on purpose, though.

So it's nice to have "own" it listed before "have" in the AMR reification table, I guess.

[flipz357]

@nschneid I find that there may be also a slight contradiction in the AMR guideline regarding ":poss / own-01 / have-03".

In the guidelines it says that In the AMR corpus, for consistency, reifications are collapsed unless needed. I interpret "unless needed" in this case as "if it's possible to apply a dereification rule according to the guidelines, then we will do so".

So while I find that this is indeed true for other relations in the reification table, it seems different with :poss:.

E.g. in bolt12_10510_9811.2 it says:

:op2 (e2 / enterprise
                              :ARG1-of (o / own-01
                                    :ARG0 (s / state)))

which, logically, if one says that reifications are collapsed per default should be:

:op2 (e2 / enterprise
                               :poss (s / state))

So even if it's technically possible, have-03 and own-01 do not seem to be collapsed to :poss in the annotated data.

[nschneid]

The example is for "state-owned enterprises"—since the ownership is expressed as a lexical verb I would expect the AMR annotators to use a frame (i.e., own-01) rather than :poss.

I interpret the guidelines as saying, in principle, every role like :poss can be viewed as syntactic sugar for a reified relation, but in practice we limit the use of reification to cases where it's necessary.

[flipz357]

Thanks, but isn't syntactic sugar to make things more concise? If we don't know whether :poss be mapped on own-01 or have-03 it's not 100% syntactic sugar anymore (in my view).

I have only wondered because of standardization in Smatch++ specifically regarding :poss vs. own-01 vs. have-03. Like if you have these three AMRs:

(e2 / enterprise
            :poss (s / state))

and

(e2 / enterprise
            :ARG1-of (o / own-01
                      :ARG0 (s / state)))

and

(e2 / enterprise
            :ARG1-of (o / have-03
                     :ARG0 (s / state)))

then, to treat parsers more fairly, Smatch++ returns the score of 1.00 for all possible pairs, since the guideline defines clear rules how to map from one to another, even though initial structure is different. Smatch would assign a different score to every pair, since it doesn't consider the reification table rules.

[nschneid]

I think it's fine for Smatch++ to treat those as equivalent. There are cases of possession where "own" would be more or less appropriate as a lexical item in English, but I think that level of nuance in terms of reification equivalence is beyond the scope of AMR.

[flipz357]

Yup, thanks for the help!