VERB -> ADJ

[uhermjakob]

One of our new summer interns, Afton Coombs, suggested to map causative verbs to "make" + adjective, e.g.

• sadden -> make sad
• simplify -> make simple

This would help to map more sentences with the same meaning to the same AMR, and make the core of word pairs such as sad/sadden semantically equivalent.

We already do not to verbalize the adjectives that causative verbs are based on:

• DO-NOT-VERBALIZE familiar TO familiarize-01
• DO-NOT-VERBALIZE liquid TO liquefy-01
• DO-NOT-VERBALIZE sad TO sadden-01
• DO-NOT-VERBALIZE simple TO simplify-01
• DO-NOT-VERBALIZE stable TO stabilize-01
• DO-NOT-VERBALIZE white TO whiten-01

VERB -> ADJ mappings would be an extension of our recently discussed NOUN -> ADJ mappings (e.g. sadness -> sad). In both cases the reverse POS mapping directions ADJ -> VERB and ADJ -> NOUN still dominate (e.g. reliance/reliant -> rely-01).

Current annotation (per guidelines) for "The girl was saddened by the disaster."

(s / sadden-01
   :ARG1 (g / girl)
   :ARG2 (d / disaster))

Proposed annotation:

(m / make-02
   :ARG0 (d / disaster)
   :ARG1 (s / sad
            :domain (g / girl)))

Scope: At least for the time being, this proposal is limited to VERB -> make-02 + ADJ. It would not include more complex mappings such as hospitalize -> admit/put/send to a hospital. I looked into how often this might occur in the corpus. Actually, a bit more often than I first thought, if we include verbs such as open, strengthen, modernize, publicize, clarify.

We would add the more common such mappings to our AMR lists so that the AMR Editor could provide guidance to annotators on specific words.

Choices include: (1) no change, (2) add to guidelines now, (3) hold off until AMR 1.5

[kevincrawfordknight]

i like the idea, though i'd rather use cause-01, since we have machinery around that already, and since make is also often (and confusingly, here) a syntactical light verb in english.

[nschneid]

This sounds like it could open up a can of worms of verbs in having PropBank frames that we ask our annotators not to use, despite being a good semantic fit. And it makes the AMR a bit more distant from the sentence/complicated to read.

If the lexical mapping is really straightforward, why not just add an automatic postprocessing step that converts from the annotated verb frames? But I suspect there will be many borderline cases that require decision-making (like knowing when to verbalize, but in reverse), putting added burden on annotators. (Does specify = make-specific? Does categorize = make-categorical? That sort of thing.)

There might be a fundamental issue here about what the "deep" category of a lexical item is: entity-like, event-like, or attribute-like? Sometimes this is apparent from derivational morphology—like sadden and sadness being variants of the attribute sad—but not always.