Revisiting focus

[nschneid]

There are some constructions where the policy for focus is not answered by the guidelines. I have been saying (and I believe we have said in the AMR Tutorial) that the main predication of the sentence is focused. Some possible exceptions:

Copular constructions

In some cases with a predicate nominal, we may want to have something other than a nominal concept as the focus:

• Deverbal argument: "He is a teacher." Should teach-01 be on top, making it equivalent to "He teaches?" Or should it be "He is a person who teaches?", with person :domain he :ARG0-of teach-01?
• Adjectives: "This is tough food to eat." food :domain this :ARG1-of [eat-01 :ARG1-of tough-01] = "This is food that is tough to eat"? Or tough :ARG1 [eat-01 :ARG1 food] = "This food is tough to eat"?
  • What about "It is tough to eat this food" (see below)?

    Information structure constructions

Focus is a property of information structure, and there are constructions that modify the default focus.

• Cleft: "It was John who ate the pizza." Places emphasis on John (as opposed to someone else). Should John be on top? Or should this be equivalent to "John ate the pizza"?
• "As for John, he ate the pizza." Same as above.

[uhermjakob]

The AMR dictionary provides this example: Mrs. O'Connor is our French teacher.

(t / teach-01 
  :ARG0 (p / person :name (n / name :op1 "Mrs." :op2 "O'Connor")) 
  :ARG1 (l / language :name (n2 / name :op1 "French")) 
  :ARG2 (w / we))

---

I would think that the following sentences would all be annotated the same:

• John ate the pizza.
• It was John who ate the pizza.
• As for John, he ate the pizza.

I think that the cleft sentence implies that the reader might very well have expected that somebody else ate the pizza instead. But placing John on top doesn't really capture that.

However, person would move to the top for:

• John, who ate the pizza
• There is a person called John who ate the pizza.

[nschneid]

OK, I should update the tutorial slides with examples of this. A summary:

1. If a predicate NP is a deverbal noun like "teacher", person is not inferred because the concept from the subject of the sentence is equated with the argument-filler (in this case, :ARG0-of teach-01). In general, types from the ontology are defaults, inferred when (i) there is a named entity without a word in the sentence to express its type (pronouns do not qualify); and (ii) a filler for a frame role is needed and no concept lexically expressed in the sentence is available.
2. Our use of the term "focus" is not the same as topic in information structure. Fronting constructions that convey topicality are normalized away for purposes of the AMR. Non-question sentences with pronominal predicate complements (e.g., "The best person for the job is him") are normalized to have the pronoun as the subject. Existentials and noun phrase "sentences" result in the main nominal concept serving as the focus of the AMR; otherwise the main predicate of the sentence (or the event it decomposes into—see #.1) will receive the focus.

[nschneid]

Remaining questions:

• Adjectives: "this is tough food to eat" (see original post)
• Questions: What is the policy for copular WH-questions with a pronominal predicate complement, like "Who is he?" and "What is that?" a. I would guess amr-unknown is the focus: amr-unknown :domain he, amr-unknown :domain that. b. But some of the instances of "Who are you?" in LPP are annotated as you :domain amr-unknown. c. Or, if we had a reification of :domain, I could see that receiving the focus.

[kevincrawfordknight]

I think we put "tough" at the top in all these cases. They all seem to be mainly saying "X is tough".

Wh-copula: Because "(l / lawyer :domain (h / he))", therefore "(a / amr-unknown :domain (h / he))". Accidentally flipping the order around ":domain" is a common annotator error -- probably that's what's happening, instead of a conscious choice about the focus.