New roles for AMR 1.5 about amr-guidelines

[kevincrawfordknight]

Good idea to supply (real) examples with suggested new roles. For ":size" --

i saw an extraordinary small person

(s / see-01
  :arg0 (i / i)
  :arg1 (p / person
          :size (s2 / small)
          :mod (e / extraordinary)))

the planet was larger than a house

(l / large
  :compared-to (h / house)
  :size-of (p / planet))

[uhermjakob]

Can you give us an idea of how narrowly or widely this would be used?

• a tall boy (size vs. height)
• a long train (size vs. length)
• XL t-shirt
• size 10.5 shoes
• 50 gallon fish tank (size vs. volume)
• 900 sq.ft. apartment (size vs. area)
• 60"x80" mattress (size vs. width and length)
• bookcase of size 60w x 15d x 68h (size vs. width/depth/height)
• 7.0 earthquake (non-spatial size)
• a big applause (non-spatial size)
• huge improvements in customer satisfaction (non-spatial size)
• small problem (non-spatial size)

[kevincrawfordknight]

i do like these instead of :size, for physical stuff:

:height
:width
:depth
:volume
:area
:length

otherwise we represent "900 sq ft apartment" with :mod

my previous examples would be :volume

:age has been very helpful to us, avoiding nasty representations of "24-year-old". :age also distinguishes between "old man" (:age) and "old friend" (:mod), supposing we are using it for "old man"

[nschneid]

But we already have special quantity concepts: volume-quantity, area-quantity, distance-quantity, etc. So wouldn't distinguishing the roles for explicit quantities be redundant?

More generally, I'm concerned that cluttering AMR with lots of (for lack of better terms) "ontology-style" as opposed to "linguistically-based" role categories will make reliable annotation much harder. What is the guiding principle for inventing roles? Why not have :birthplace, :income, :mother-in-law, :favorite-color?

I like how we currently rely on concepts/frames for most of the specialized parts of the meaning, such that (a) AMR roles are necessary only for relations that cannot be suitably represented with core arguments, and (b) the granularity of roles aims to be broad—though sometimes it feels too broad, as with :mod and formerly :poss. I would state the guiding principle as: favor verb/adjective/noun frames where there is some word in the sentence whose lexical semantics is rich enough to convey the relation. And, on occasion, fall back to agreed-upon frames for meanings expressed in some other way (as we do with say-01 and opine-01).

:age has been very helpful to us, avoiding nasty representations of "24-year-old". :age also distinguishes between "old man" (:age) and "old friend" (:mod), supposing we are using it for "old man"

The main argument for :age IMO is that it is both frequent and expressed with strange syntax (such as the N years old or N years of age constructions) such that syntax-based frames would be unsatisfying. But I'd actually be in favor of always reifying it to age-01 because it is such a "deep" semantic role.

[kevincrawfordknight]

Right, a thorough study of :mod in our existing AMRs will shed some light on this. A guiding principle is to root out syntactic relations and strive to assign the same AMR to paraphrases.

[uhermjakob]

A large number of new roles would probably also lead to an equally large number of new reification frames. The sentence The house is big. would presumably then have to be annotated as

(h / have-volume-91
   :ARG0 (h2 / house)
   :ARG1 (b / big))