Constructions: An exploration in PropBank & AMR

[cbonial]

Yesterday, we were able to discuss constructions with Kevin during his visit to CU. This document covers some theoretical linguistic background to constructions, as well as their treatment in PropBank and AMR. Finally, it includes proposals for how certain constructions could be treated more consistently with a deeper representation in AMR. Please don't be intimidated by the length -- it's a choose your own adventure! Skip to the actual proposals if desired, and please note that much of the length is just AMR annotation examples. Here's a link to the document:

https://drive.google.com/file/d/0B-54bBqKapwXQWJfdDUycjlmdlk/view?usp=sharing

[nschneid]

This is an important discussion to start having. Thanks for the detailed writeup!

I'm sure we'll want to talk about this on the call, but my initial reaction is as follows:

Constructions are (according to the theory) mappings between form and meaning. Because AMR aims to abstract away from syntactic form, IMO it would be more in the spirit of AMR to express the meaning with non-core roles, or with lexical predicates (e.g., result-01 for resultatives if we want to distinguish types of causation).

E.g., for He blinked the snow off of his eyelashes, how about:

(b / blink-01
      :ARG0 (h / he)
      :manner-of (m / move-01
            :ARG0 h
            :ARG1 (s / snow)
            :source (e / eyelash
                  :part-of h)))

which would be the same as for He moved the snow from his eyelashes by blinking. (I would use :means-of instead of :manner-of if AMR made the distinction.)

If we were to start adopting construction frames directly into AMR, it seems like it would become hard to characterize the set of things that AMR does abstract away from. We'd effectively need to say there is a list of syntactic phenomena that AMR handles in a "deep" way, and there is a list of other syntactic phenomena that it handles in a shallower way (construction frames).

If you think about it, thus far, we've really been addressing constructions in the documentation, but keeping them out of the representation itself. (Arguably, the main reason for the AMR Dictionary is that there are so many dang constructions to deal with!) Assuming that PropBank goes down the road of annotating constructions explicitly, it might make sense to formalize these mappings from construction frames to AMR fragments so the conversion can be automated.

[uhermjakob]

I like Nathan's AMR. The sentence reminds me of the several instances of argument raising that we recently came across (again).

(1) I don't think that the Yankees will win. (win-01 :polarity -) (2) John seems to be sleeping. (sleep-01 :ARG0 John) (3) The weapons are intended for export. (export-01 :ARG1 weapon) (4) He blinked the snow off of his eyelashes. (move-01 :ARG1 snow :source eyelash)

All sentences have raised arguments, but in the last sentence the embedded event that the arguments were raised from is only implied.

[kevincrawfordknight]

Claire's constructions and examples are great. These are things that are handled inconsistently by AMR annotators.

(Although, I was really amazed that our annotators took "The longer he is around, the more miserable I will be" and turned it into ":condition"! As in "if he is around longer, I will be more miserable". Any quibbles aside, that is an excellent use of AMR primitives.)

As for what representations to adopt, I agree that we don't want English syntax in AMR. There may be multiple constructions for expressing the same basic meaning.

For most of Claire's constructions, we can use or invent an AMR role (:cause, :condition, etc). We can then use reified versions if we want to modify the relation, or promote it to the root of a substructure.

"Blink the snow off" is the only one where you feel like additionally introducing an un-mentioned concept ("move-01") into the AMR.

Then, it's a matter of documenting all these things.

[kevincrawfordknight]

Just a digression... page 22 has an interesting idea that doesn't necessarily have anything to do with constructions, that of having a degree frame. I remember us stressing a bit about:

(X / tall :arg0 W :degree Y :compared-to Z)

because ":degree" and ":compared-to" were somehow bound tightly to each other in our minds, but only loosely parallel in the AMR. Having a degree frame would enable us to wrap up :compared-to, :consequence, and possible other things into a single structure. Had we thought of this at the time, we might have adopted it.

[uhermjakob]

So, yes, our current :degree/:compared-to notation is a bit sloppy and syntacticky, but as we found in the past, often much more annotator-friendly than "pure" semantic annotations such as:

(1) Comparatives "Ulf is taller than Claire."

(e / exceed-01
  :ARG0 (d / distance-quantity
          :quant-of (t / tall
                      :domain (p / person :name (n / name :op1 "Ulf"))))
  :ARG1 (d2 / distance-quantity
           :quant-of (t2 / tall
                        :domain (p2 / person :name (n2 / name :op1 "Claire")))))

(2) Clauses of proportion "The more a product costs, the less people will buy it."

(m / monotonic-decrease-91
  :ARG0 (m2 / monetary-quantity
           :ARG2-of (c / cost-01
                      :ARG1 p
                      :ARG3 p2))
  :ARG1 (n / number
          :quant-of (p / product
                      :ARG1-of (b / buy-01
                                 :ARG0 (p2 / person)))))

monotonic-decrease-91 is inspired by http://en.wikipedia.org/wiki/Monotonic_function.

[nschneid]

Since Claire's have-degree-91 proposal would replace :compared-to, here are examples of prepositions that we supersense-tag as Comparison/Contrast:

a drawing after Millet's The Reapers.; they named her Pauline, after Barbara's mother.; the benefits must be weighed against the costs.; His choice of Indonesia , as opposed to the United Kingdom or the USA , as his first foreign destination was controversial; a radial as opposed to a concentric circular structure beside Paula she always felt clumsy.; the difference between income and expenditure.; these fees are quite distinct from expenses.; he used to have a car like mine; they were like brothers; she looked nothing like Audrey Hepburn.; he was screaming like a banshee.; students were angry at being treated like children.; Our council tax is much higher than in larger towns; It turned out to be better in looks than taste; everything is worse than it was; He likes getting in on the act too and has appeared in more productions than he cares to count.; On one occasion a female passenger leaned on him for more than just a friendly chat. They just seem to do what they are told rather than have their own personal opinion.; we would rather be able to see than be blind. the club's nothing to what it once was.;

Some of these would be handled in AMR with PropBank args. Would the rest be covered by have-degree-91 and resemble-01?

[cbonial]

I will not be able to call in to the AMR meeting 3/30 as I'm attending the NSF workshop on unified annotation tooling. I will provide examples of compared-to for discussion, and add some information on guidelines for motion constructions, for the next meeting, 4/6.

[cbonial]

Please find the proposal for the reification of :degree, have-degree-91, below. I've included examples of how this could replace the current combination of :degree and :compared-to, as well as a couple of examples involving the degree-consequence construction. There's a brief note on the feasibility of converting some of these automatically as well, but this will probably require some discussion with Ulf to determine.

Reification of Degree instead of having multiple rolesets for constructions relating to degrees to which an attribute applies to a entity, create the reification for :degree, have-degree-91.

Have-degree-91 Arg1: attribute (e.g. tall) Arg2: domain, entity characterized by attribute (e.g. girl) Arg3: degree itself (e.g. more, less, equal) Arg4: compared-to (e.g. (than the) BOY) Arg5: consequence, result of degree (e.g. (not tall enough) TO RIDE THE ROLLERCOASTER)

Examples with compared-to [1]   consensus   isi_0001.27   (snt. 26 in workset  guidelines, last updated on Mon May 14, 2012) the girl is taller than the boy Current: (t / tall       :degree (m / more)       :domain (g / girl)       :compared-to (b / boy))

Proposed: (t / tall       :ARG1-of (h / have-degree-91             :ARG2 (g / girl)             :ARG3 (m / more)             :ARG4 (b / boy)))

[4]   consensus   isi_0002.256   (snt. 293 in workset  guidelines, last updated on Wed Jul 31, 2013) Texas is the second largest state in the U.S. Current: (s / state       :mod (l / large             :degree (m / most                   :ord (o / ordinal-entity :value 2))             :compared-to (c / country :wiki "United_States"                   :name (n2 / name :op1 "U.S.")))       :domain (s2 / state :wiki "Texas"             :name (n / name :op1 “Texas")))

Proposed: (s / state       :mod (l / large             :ARG1-of (h / have-degree-91                   :ARG2 s                   :ARG3 (m / most                         :ord (o / ordinal-entity :value 2))                   :ARG4 (c / country :wiki "United_States" :name (n2 / name :op1 "U.S."))))       :domain (s2 / state :wiki "Texas" :name (n / name :op1 "Texas")))

[8]   consensus   isi_0002.571   (snt. 608 in workset  guidelines, last updated on Thu Nov 6, 2014) He runs as fast as lightning. Current: (r / run-02       :ARG0 (h / he)       :manner (f / fast             :degree (e / equal)             :compared-to (l / lightning)))

Proposed: (r / run-02       :ARG0 (h / he)       :manner (f / fast             :ARG1-of (h2 / have-degree-91                   :ARG2 h                   :ARG3 (e / equal)                   :ARG4 (l / lightning))))

Plausible to do automatically? General heuristics Thing modified -> Arg2-of have-degree-91 (hardest to find? Can be :domain arg, or thing with :mod, and [4] demonstrates that these can co-occur) Thing currently modified by :degree -> Arg1-of have-degree-91 :Degree argument -> Arg3-of have-degree-91 :Compared-to argument -> Arg4-of have-degree-91

Examples with the Degree-Consequence Construction

[2]   cjconsensus   DF-199-192794-661_2329.5   (snt. 20 in workset  dfa-wset-6, last updated on Tue Sep 3, 2013) Now the wrist units at the local drugstores come with a read-outbox that is too wide for my 5 1/2 inch wrist. Current: (c / come-01       :ARG1 (u / unit             :mod (w / wrist)             :location (d / drugstore                   :mod (l / local)))       :time (n / now)       :accompanier (o / outbox             :mod (w2 / wide                   :degree (t / too)                   :prep-for (w3 / wrist                         :part-of (i / i)                         :mod (d2 / distance-quantity :quant 5.5                               :unit (i2 / inch))))             :instrument-of (r / read-01)))

Proposed (note the introduction of concept “fit” with polarity -): (c / come-01       :ARG1 (u / unit             :mod (w / wrist)             :location (d / drugstore                   :mod (l / local)))       :time (n / now)       :accompanier (o / outbox             :mod (w2 / wide                   :ARG1-of (h / have-degree-91                         :ARG2 o                         :ARG3 (t / too)                         :ARG5 (f / fit-01 :polarity -                               :ARG1 o                               :ARG2 (w3 / wrist                                     :part-of (i / i)                                     :mod (d2 / distance-quantity :quant 5.5                                           :unit (i2 / inch))))))             :instrument-of (r / read-01)))

[1]   cjconsensus   PROXY_AFP_ENG_20080813_1032.33   (snt. 55 in workset  proxy-19-wset-69, last updated on Wed Jul 24, 2013) It is too early to reach any conclusion as to the motive and identity of the attackers. Current: (e / early       :degree (t / too)       :domain (c2 / conclude-01             :ARG1 (a / and                   :op1 (m / motive                         :poss (p / person                               :ARG0-of (a2 / attack-01)))                   :op2 (i / identity                         :poss p))))

Proposed (note the introduction of polarity - on “conclude”): (e / early       :ARG1-of (h / have-degree-91             :ARG3 (t / too)             :ARG5 (c2 / conclude-01 :polarity -                   :ARG1 (a / and                         :op1 (m / motive                               :poss (p / person                                     :ARG0-of (a2 / attack-01)))                         :op2 (i / identity                               :poss p)))))

[cbonial]

Concerning Nathan's previous comment, here are my (off-the-cuff) thoughts:

1. a drawing after Millet's The Reapers.;
   • resemble.01
2. they named her Pauline, after Barbara's mother.;
   • name.02 :arg2 namesake
3. the benefits must be weighed against the costs.;
   • weight.04, arg2 (consider) in comparison to
4. His choice of Indonesia , as opposed to the United Kingdom or the USA , as his first foreign destination was controversial;
   • instead-of-91
5. a radial as opposed to a concentric circular structure
   • instead-of-91
6. beside Paula she always felt clumsy.;
   • :compared-to Paula-> :arg4-of have-degree-91
7. the difference between income and expenditure.;
   • differ.02, arg1, arg2 thing that is different, (different) from what
8. these fees are quite distinct from expenses.;
   • distinguish.01 args? Don’t have access to find this alias right now
9. he used to have a car like mine; they were like brothers; she looked nothing like Audrey Hepburn.; he was screaming like a banshee.; students were angry at being treated like children.;
   • all resemble.01
10. Our council tax is much higher than in larger towns; It turned out to be better in looks than taste; everything is worse than it was; He likes getting in on the act too and has appeared in more productions than he cares to count.; On one occasion a female passenger leaned on him for more than just a friendly chat.; the club's nothing to what it once was.;
    • all :compared-to -> :arg4-of have-degree-91
11. we would rather be able to see than be blind. They just seem to do what they are told rather than have their own personal opinion.;
    • all instead-of-91

[nschneid]

The examples look good as far as I can tell.

What about "good enough for me"? Does have-degree-91 need a "judge" role?

[cbonial]

Please find proposed general guidelines on constructions in AMR here:

https://drive.google.com/file/d/0B-54bBqKapwXV0hpWWNCT3lUMEU/view?usp=sharing

[cbonial]

In our last call, we discussed creating a "challenge set" of sentences, starting with those that are marked as constructions in PB. I have compiled those sentences, and separated them by...

1. Degree-Consequence Constructions a. that require no hallucination b. that require some hallucination c. that have minimal cueing, besides the degree word
2. The Xer, The Yer Constructions
3. Comparative Constructions (in case this is helpful in determining whether we want to adopt the proposed have-degree-91).

Please find the full list here, and we can consider which to AMR: https://drive.google.com/file/d/0B-54bBqKapwXRUZaUWNaUzY4OW8/view?usp=sharing

You'll see that there are a couple of annotation mistakes in this list. According to the PB guidelines, constructions that implicitly invoke Comparatives or Degree-Consequence constructions should only be marked with CXN if arguments licensed by this construction (outside of just the degree word itself) are present. If they are not present, annotators are simply to mark the degree word as :extent (we have no :degree in PB). So the annotation mistakes reflect some confusion on this point, and this is an issue that we need to consider carefully for AMR -- do we want to allow annotators to use have-degree-91 when the construction is implicit, but no more than a degree word is present? Or do we want them to have the ability (or be required to) use simple :degree in these cases? I'm hoping that we can find a way to implement have-degree-91 in such a way that :degree X and ArgX-of have-degree 91 are equivalent, so that annotators can use either option.

[cbonial]

The last issue to resolve concerning :degree is the Xer the Yer construction. In one of our last discussions of this in comparison to other degree constructions, we concluded that this construction should have its own roleset to avoid confusing overlap. Ulf has previously proposed the following roleset:

I think folks were in support of the idea behind this roleset, but objected to the potentially intimidating label. A proposal for discussion may be something like:

proportional-change-91 Arg1 - 1st item changing Arg2 - 2nd item changing in some proportion to Arg1

So in Ulf's example above, m2 would be Arg1, while n would be Arg2 (I don't think it should have an Arg0 since there's nothing agent-like). The longer I think about this, the harder it is to think of good descriptions, so please make suggestions!

[nschneid]

From Fillmore et al. 1988: "This structure is used for expressing a correlation between an independent variable and a dependent variable." Can we just reuse correlate-01?

To propose something a bit shallower for "The more a product costs, the less people will buy it":

(c / correlate-01
    :ARG1 (c2 / cost-01
        :ARG1 (p / product)
        :ARG3 p2
        :degree (m / more))
    :ARG2 (b / buy-01
        :ARG0 (p2 / person)
        :ARG1 p
        :degree (l / less)))

I.e., a greater extent of cost correlates with a lesser extent of buying.

I am sympathetic to the desire to go deeper to the amount of money and the quantity of goods, but fear it would require too much hallucination to be practicable.

[cbonial]

Links for the proposed guideline addition for constructions generally (including information on Have-Degree-91) and proposed additions for examples of have-degree-91 for the dictionary can be found below.

Construction Guidelines: https://docs.google.com/document/d/1tsUjoImZq-GLyIyCG3EZsOP9URFwaMOomSPFyMMwFwo/edit?usp=sharing

Dictionary Examples: https://docs.google.com/document/d/1ISCnY8widSaWXXSOq0mvGJ21R_k7jJL0IuPllJ16_gE/edit?usp=sharing

The final challenge of this topic was coming up with an appropriate roleset for "The X-er the Yer" construction. As I began playing around with creating AMRs for this construction, I felt more and more strongly that Nathan was on the right track by associating this with the concept of correlation. If we restrict ourselves to a roleset called "the-the" or "the-x-the-y" (remember that the editor does not allow for camel-casing or caps), I worry that the resulting roleset will be totally opaque to our users. However, I can also understand the desire to keep this construction separate from other realizations of correlate-01 more generally. So I'm proposing a new roleset, correlate-91, that is defined as a specific type of correlation expressed by the X-er the Y-er construction. Thus, it will be an identifying marker for this construction, but it still carries some understandable semantic content for our users. I feel like generalizing to some type of correlation semantics (instead of defining this solely in terms of the English construction) is also in the spirit of AMR because it highlights semantics over syntax.

With final approval of the correlate-91 roleset, I will add the text of the proposed Construction guidelines to our guidelines as a new section under Part III.

[nschneid]

Excellent! A few questions:

1. Can you add an example for quantity comparisons, e.g. "more apples than pears" or "He bought as much beer as he could carry."? I'm guessing the plan is to continue to hallucinate quantity for these.
2. What about when events are being compared on a scale? "And the city decided to treat its guests more like royalty or rock stars than factory owners ."—maybe this would have resemble-01 as an argument of have-degree-91?
3. Will comparative as...as use have-degree-91 with :ARG3 (e / equal)?
4. Will :compared-to be removed entirely from AMR?

[cbonial]

Sorry for the delay in responding to this, I've moved across the country to a new job at ARL in MD! I'm still settling in, but hope to continue contributing to and participating in AMR discussions.

I've updated the document with examples addressing (1) and (2) that Nathan pointed out above: https://docs.google.com/document/d/1ISCnY8widSaWXXSOq0mvGJ21R_k7jJL0IuPllJ16_gE/edit?usp=sharing

Concerning (2), I'm not sure I agree with the consensus annotation of this sentence, but I tried to stick to it as closely as possible, so I did NOT introduce resemble-01 since the original annotation did not have this. Looking through examples of "resemble" used for usages containing "like," I noticed that these do not seem to include verbs that use a numbered argument for the an attribute or depictive (which treat-01 does), so I suppose that is the logic behind not using it here?

Thus, I added another existing example that does use "resemble" for "like" in the proposed dictionary examples, but tweaked it slightly so that it also has a comparison: "They were drooling more like schoolboys than Fortune 500 Executives." Please have a look at this example and let me know if there are any objections-- essentially I tried to capture that "their drooling resembles the drooling of schoolboys more than executives," hence there's a slight temptation to add a third "drooling," (executive drooling?). However, I think dropping it is analogous conceptually to the decision to drop a second mention of, for example, "tall" when essentially saying "the girl's tallness is more than the boy's (tallness)" in a simple comparison like "the girl is taller than the boy."

Concerning (3): Yes, comparatives as...as should use have-degree-91 with degree equal, and there's an example of this in the dictionary examples document linked above.

(4): Yes, the idea of the proposal was that the comparative construction, which previously evoked :degree and :compared-to, will now evoke have-degree-91. Although simple :degree could still be used (the idea being that have-degree-91 is the reification), a construction that makes use of other arguments should invoke the have-degree-91 relation.

[cbonial]

In response to discussion in our last call regarding sentences of the type (Verb more like X than Y, e.g. They dance more like natives here than like normal people), I've written up some thoughts on this topic: https://docs.google.com/document/d/1qXNhsoM0JEF3Vts4w3oF_6cn3iy8BLnt9_QF7pn4jxo/edit?usp=sharing.

However, I've also asked Tim to pass the have-degree-91 roleset, a couple of examples, and this sentence on to a couple of annotators to see what they do with it. I think this will inform how we should move forward given that the roleset is certainly clunky, but possibly adequate, for these admittedly rare usages.

[nschneid]

Regarding dropping the second resemble-01:

it’s striking how much less readable this is than the case of “tallness:” is it clear that we are comparing two resemblances?

I agree, it seems unclear without a resemble-01 in the :ARG4. Perhaps the convention should be that in the absence of an :ARG2, there must be a fully fleshed-out :ARG4?

what happens if we introduce another, clearer attribute?

I wonder if the resemblance in the AMR should be between the two dancing events, not specifically between their manners. Then the problem goes away, right? "They dance wildly, more like the natives here than like normal people." means 'They dance wildly, and their dancing resembles the natives' dancing more than it resembles normal people's dancing.' (With no implication of degrees of wildness, or that the standards of comparison are wild.)

[cbonial]

Update on treatment for constructions like "They dance more like the natives here than normal people." Here's what we've done (thanks to CU folks for the help):

1. 2 CU annotators created AMRs for this sentence using have-degree-91. This reflected confusion over what is arg1 (attribute) and arg2 (domain, thing with attribute), indicating that these two arguments are not clearly distinct within these constructions.
2. Tim and I created new candidate rolesets specifically for this construction, and we (Tim, Martha and I) voted on our favorite.
3. The resulting roleset: Have-degree-of-resemblance-92 Arg1: Thing resembling things Arg2: First resemblance under comparison Arg3: Second resemblance under comparison Arg4: degree word comparing Arg2 to arg3
4. And example annotation for sentence "They dance more like the natives here than normal people."

1. Thoughts?

[nschneid]

:+1:

[cbonial]

The final, final addition to the guidelines on constructions (with new portion on the constructions comparing resemblances) and relevant dictionary examples for have-degree-91 and related constructions can be found here:

https://drive.google.com/file/d/0B-54bBqKapwXTXNLTVBQSWRVSjA/view?usp=sharing https://drive.google.com/file/d/0B-54bBqKapwXamJLYlhrWDE5MFU/view?usp=sharing

I am happy to incorporate the guideline addition whenever that is appropriate, just let me know. The examples should be added to the Editor dictionary once the new treatment of comparisons is adopted.

[nschneid]

@ida-szubert points out that the adopted interpretation of have-degree-91, though sensible for comparatives, is a bit strained for superlatives:

[4] consensus isi_0002.256 (snt. 293 in workset guidelines, last updated on Wed Jul 31, 2013) Texas is the second largest state in the U.S.

(h / have-degree-91
      :ARG1 (l / large)
      :ARG2 (s / state
            :domain (s2 / state :wiki "Texas" :name (n / name :op1 "Texas")))
      :ARG3 (m / most
            :ord (o / ordinal-entity :value 2))
      :ARG4 (c / country :wiki "United_States" :name (n2 / name :op1 "U.S.")))

This is strange because the reference point of comparison should really be a set, not a single entity. The above suggests that Texas is being compared to the U.S., whereas it is really being compared to other states in the U.S. (A literal reading: "Compared to the United States, the state of Texas is the second largest state.")

Why not move the explicit mention of "state" (s) so it is part of the reference point:

(h / have-degree-91
      :ARG1 (l / large)
      :ARG2 (s2 / state :wiki "Texas" :name (n / name :op1 "Texas"))
      :ARG3 (m / most
            :ord (o / ordinal-entity :value 2))
      :ARG4 (s / state 
           :part-of (c / country :wiki "United_States" :name (n2 / name :op1 "U.S."))))

I.e.: compared to the [other] states of the U.S. / out of all the states in the U.S., Texas is the second largest.

[nschneid]

(Swapping :ARG1 and :ARG2, as agreed in today's meeting:)

the tallest girl = the girl who is (the) tallest

(g / girl 
    :mod (t / tall 
          :degree (m / most)))

the tallest of the girls

(g / girl
    :ARG1-of (h / have-degree-91
        :ARG2 (t / tall)
        :ARG3 (m / most)
        :ARG4 (g2 / girl)))

the tallest girl on the team = the tallest of the girls on the team = the tallest girl of the girls on the team = the girl who is the tallest (of the girls) on the team

(g / girl 
    :ARG1-of (h / have-degree-91
        :ARG2 (t / tall)
        :ARG3 (m / most)
        :ARG4 (g2 / girl
            :part-of (t2 / team))))

the tallest girl of the people on the team = the girl who is the tallest of the people on the team

(g / girl 
    :ARG1-of (h / have-degree-91
        :ARG2 (t / tall)
        :ARG3 (m / most)
        :ARG4 (p / person
            :part-of (t2 / team))))

[nschneid]

A tricker case is when the degree-modified property is an argument-taking frame: a sentence in the data mentions "the easiest opponent for obama to defeat in the general election".

Let's consider "the easiest person to defeat". I assume that would simply be:

(p / person
      :ARG1-of (d / defeat-01
            :ARG1-of (e / easy-05
                  :degree (m / most))))

But what if it's "the easiest of the candidates to defeat"? Do we say have-org-role-91 :ARG2 [easy-05 :ARG1 [defeat-01 :ARG1 candidate]]? If so, is the candidate variable repeated as the :ARG1? One possibility:

(c / candidate
      :ARG1-of (d / defeat-01
            :ARG1-of (e / easy-05
                  :ARG2-of (h / have-degree-91
                        :ARG1 c
                        :ARG3 (m / most)
                        :ARG4 (c2 / candidate)))))

We really should be comparing the ease of defeating candidate c vs. the ease of defeating other candidates c2. So an alternative could be:

(c / candidate
      :ARG1-of (d / defeat-01
            :ARG1-of (e / easy-05
                  :ARG1-of (h / have-degree-91
                        :ARG3 (m / most)
                        :ARG4 (e2 / easy-05
                              :ARG1 (d2 / defeat-01
                                    :ARG1 (c2 / candidate)))))))

without an :ARG2. This has an elegant symmetry about it, but may be a bit too deep to be practical.

[nschneid]

The [fund's average] maturity is the longest since August. (adapted from wsj_0004.4, where the :time is under long-03 in the current AMR). Proposed:

(h / have-degree-91
    :ARG1 (m / maturity)
    :ARG2 (l / long-03
        :ARG1 m)
    :ARG3 (m2 / most)
    :ARG4 (m3 / maturity
        :time (s / since
            :op1 (d4 / date-entity :month 8))))

He is easily among the most competent politicians we have. — currently:

(i / include-91 
      :ARG1 (h / he) 
      :ARG2 (p / politician 
            :mod (c / competency 
                  :degree (m / most)) 
            :ARG1-of (h2 / have-03 
                  :ARG0 (w / we))) 
      :ARG1-of (e / easy-05)) 

I think this would remain unchanged.

[cbonial]

I like the notion of making the set/superset more explicit (e.g. specifying that the comparison is to GIRLS on the team or to STATES in the U.S. as opposed to comparing to team or country). I think that this: "We really should be comparing the ease of defeating candidate c vs. the ease of defeating other candidates c2" takes it a step further, making explicit that we're comparing TALLNESS of girls and LARGENESS of states to continue the previous examples from our discussion. This does end up leading to a lot of repetition. If we think that all comparisons automatically invoke a set/superset with a shared property under comparison, perhaps none of it is necessary to make explicit and that was the assumption before. My conclusion is that going to the level of GIRLS/STATES as opposed to the property itself is a nice, happy medium.

[nschneid]

@cbonial: OK, I think that's a nice principle. How would you handle "the easiest of the candidates to defeat"?

[uhermjakob]

Consensus for "The candidate that is easiest to defeat."

(c / candidate
      :ARG1-of (d / defeat-01
            :ARG1-of (h / have-degree-91
                  :ARG2 (e / easy-05
                        :ARG1 d)
                  :ARG3 (m / most)
                  :ARG4 (d2 / defeat-01
                        :ARG1 (c2 / candidate)))))

"This is the easiest candidate to defeat." = "This candidate is easiest to defeat.":

(h / have-degree-91
      :ARG1 (c / candidate
            :ARG1-of (d / defeat-01)
            :mod (t / this))
      :ARG2 (e / easy-05
            :ARG1 d)
      :ARG3 (m / most)
      :ARG4 (d2 / defeat-01
            :ARG1 (c2 / candidate)))

[uhermjakob]

All participants on the AMR phone on July 18, 2016, love Nathan's examples on June 27, 2016 ("girls").

"The tallest new girl on the team." = "The tallest of the new girls on the team."

(g / girl
      :ARG1-of (h / have-degree-91
            :ARG2 (t / tall)
            :ARG3 (m / most)
            :ARG4 (g2 / girl
                  :ARG1-of (n2 / new-02
                        :ARG2 t2)
                  :ARG0-of (h2 / have-org-role-91
                        :ARG1 t2)))
      :ARG1-of (n / new-02
            :ARG2 (t2 / team))
      :ARG0-of (h3 / have-org-role-91
            :ARG1 t2))

Nathan: Wonder whether we should have a special frame for superlatives. Let's investigate whether include-91 might help to overcome redundancies.

[kiragrif]

Here is what I would propose for include-91:

Sentence: She is the tallest new girl on the team
(i2 / include-91
      :ARG1 (s / she
            :mod (t / tall
                  :degree (m / most)))
      :ARG2 (g2 / girl
            :ARG1-of (n / new-02
                  :ARG2 (t2 / team))))

[timjogorman]

Could we just do:

(a / and
      :op1 (h / have-degree-91
            :ARG1 (g / girl)
            :ARG2 (t / tall)
            :ARG3 (m / most)
            :ARG4 g2)
      :op2 (i2 / include-91
            :ARG1 g
            :ARG2 (g2 / girl
                  :ARG1-of (n2 / new-02
                        :ARG2 t2)
                  :ARG0-of (h2 / have-org-role-91
                        :ARG1 t2))))

[nschneid]

@timjogorman suggested that some sentences can be borderline between comparative and superlative; we should decide whether to AMR them the same. E.g.: "She is taller than all the other girls on the team = She is the tallest girl on the team"

Another example to work out: "She is the tallest girl on the team and the youngest kid in the class" (multiple superlatives).

[nschneid]

Conclusion from today's meeting: introduce :ARG5 which is the containing set for superlatives (removing the need for explicit include-91). The consequence role will be bumped up to :ARG6.

Thus:

She is the tallest new girl in the room. 
= She is the tallest of the new girls (who are) in the room.

(h / have-degree-91
    :ARG1 (s / she)
    :ARG2 (t / tall)
    :ARG3 (m / most)
    :ARG5 (g / girl
        :ARG1-of (n / new-02)
        :location (r / room)))

She is the tallest girl in the room and the youngest kid in the building. 

(a / and
    :op1 (h / have-degree-91
        :ARG1 (s / she)
        :ARG2 (t / tall)
        :ARG3 (m / most)
        :ARG5 (g / girl
            :location (r / room)))
    :op2 (h / have-degree-91
        :ARG1 s
        :ARG2 (y / young)
        :ARG3 (m2 / most)
        :ARG5 (k / kid
            :location (b / building))))

This is the easiest candidate to defeat. 
= This candidate is easiest to defeat. 
= Of the candidates, this one is easiest to defeat.

(h / have-degree-91
      :ARG1 (c / candidate
            :mod (t / this))
      :ARG2 (e / easy-05
            :ARG1 (d / defeat-01 :ARG1 c))
      :ARG3 (m / most)
      :ARG5 (c2 / candidate))

("She is the tallest girl on the team" would presumably require have-org-role-91.)

[cbonial]

I've updated both the guidelines for handling constructions and the dictionary topic examples, explanation for have-degree-91 and related topics (have-quant-91, correlate-91, etc). They now include an arg5 for supersets in superlatives and the consequence/result argument has been shifted to arg6. For anyone who has time, I would appreciate another set of eyes on these documents, suggestions for other examples to include and/or updates to wording that would clarify the guidelines.

AMR_Constructions_ProposedGuidelines_updatedAug2016.docx DictExamples_HaveDegree_haveQuant_updatedAug2016.docx

[nschneid]

Looks good! A couple of small things:

AMR_Constructions_ProposedGuidelines_updatedAug2016.docx

• Below example (12) it is noted that have-degree-91 should often be the focus. Why isn't it the focus in (12) and (13)?

DictExamples_HaveDegree_haveQuant_updatedAug2016.docx

• The ARG6 (Degree Consequence construction) has different AMRs in the two documents. In the first doc, (13) "It is too early to reach any conclusion" has an ARG6 of conclude-01 :polarity -, while the same sentence in the second doc has an ARG6 of possible :polarity - :domain conclude-01. Should the annotator always write an ARG6 that can be prefaced with "therefore"? Or should we expect that the negative polarity of the consequence can be inferred from the use of "too" (rather than "enough")?
• Typo: gradeable -> gradable; and missing space before the last word of the same paragraph.

[cbonial]

Thanks, Nathan! I've updated...

• the examples you pointed out so that have-degree-91 is the root node
• eliminated the "possible" in the arg6 of the degree-consequence example (I went back and forth on this, but I don't think it's necessarily licensed here and I don't think we want annotators tempted to introduce it all the time)
• some of the language in notes for degree-consequence that makes it clear that arg6 can generally be introduced with "therefore" and that polarity depends on context, but that constructions with "too" likely will introduce negative polarity.
• fixed typos and numbering, some inconsistencies in formatting and overall organization.

Here are the new docs: AMR_Constructions_ProposedGuidelines_updatedAug2016.docx DictExamples_HaveDegree_haveQuant_updatedAug2016.docx