Guidelines and annotation docs have different lists of reified concepts

goodmami commented 5 years ago

The list of reified concepts at https://www.isi.edu/~ulf/amr/lib/roles.html has 7 relation-concept pairs that are not in the list at https://github.com/amrisi/amr-guidelines/blob/master/amr.md#reification:

Relation	Concept
`:cost`	`cost-01`
`:employed-by`	`have-org-role-91`
`:li`	`have-li-91`
`:meaning`	`mean-01`
`:ord`	`have-ord-91`
`:role`	`have-org-role-91`
`:superset`	`include-91`

There is also one pair in the guidelines that's not in the docs:

Relation	Concept
`:value`	`have-value-91`

Should we consider the union of the two lists as the full set?

edit: removed :time and :topic from the first table as they do in fact exist in both places

nschneid commented 5 years ago

@ulfulf can confirm but I believe the annotation docs are more up-to-date than the guidelines.

have-value-91 occurs in only one release AMR (in the BioAMR corpus).

uhermjakob commented 5 years ago

Thanks for pointing this out. Some of these pairs are valid role/reification pairs, some "roles" are only shortcuts.

Valid AMR role/reification pairs:

:li have-li-91
:ord have-ord-91
:time be-temporally-at-91
:topic concern-02
:value have-value-91 (exceedingly rare; just added to https://www.isi.edu/~ulf/amr/lib/roles.html)

Not valid AMR roles; used only as shortcuts in the AMR Editor:

:cause cause-01
:cost cost-01
:employed-by have-org-role-91
:meaning mean-01
:role have-org-role-91
:subset include-91
:superset include-91

Note: The shortcuts' purpose is to make life easier for human AMR annotators. The AMR Editor automatically and instantly expands these shortcuts from pseudo-roles to their AMR reifications. So the shortcuts will not appear in the AMR corpus.

In the table at the bottom of https://www.isi.edu/~ulf/amr/lib/roles.html these were marked as shortcuts by being colored grey and indicating the "shortcut" nature upon mouse-over. Probably not obvious enough, so I now explicitly mark them as "shortcut only".

Updated: https://github.com/amrisi/amr-guidelines/blob/master/amr.md, https://www.isi.edu/~ulf/amr/lib/roles.html

goodmami commented 5 years ago

Great, thank you for clarifying and for updating the docs and guidelines. While we're on the subject, here are a few more things I noticed (after originally filing this issue):

It appears that the arguments of :accompanier are backwards (should be domain=ARG1 and range=ARG0, I think). I only checked the argument order of a few roles, so there may be other occurrences.
:beneficiary has two reifications in the docs (benefit-01 and receive-01) but only one in the guidelines
While many(roles)-to-one(concept) mappings seem unambiguous because they select different arguments (such as :subset and :superset), the one-to-many mappings (:beneficiary and :poss) are ambiguous. In other words, if the two concepts (benefit-01/receive-01 and have-03/own-01) have different meanings, then there is information loss when collapsed to a relation and information needed for reification. Are there guidelines for dealing with these, or is it just up to the annotators intuition about which to use?

goodmami commented 7 months ago

4.5 years later... My question above about :accompanier was resolved in #269 along with :example. Similarly, :poss was addressed in #262.

I'm not sure about the receive-01 reification of :beneficiary. I've been mapping it with domain=ARG2 and range=ARG0 following this frame: https://www.isi.edu/~ulf/amr/ontonotes-4.0-frames/receive-v.html.

I'll close this issue.

amrisi / amr-guidelines

Guidelines and annotation docs have different lists of reified concepts #239