amrisi / amr-guidelines

246 stars 87 forks source link

earliest, latest #225

Closed cbonial closed 6 years ago

cbonial commented 6 years ago

It's clear to me from the updated guidelines what to do with earlier, later, using :time before or after, but it's not clear to me what to do with earliest? latest? The only consensus example is a little bit confusing because it's sort of an MWE: "next week at the earliest". Here, the distinction between earlier/earliest is ignored, and this is also treated as :time before... It makes sense in this example ("not expected to act before next week"), but I'm not sure how this should generalize to other usages, like "this is the earliest morphological sign of..."

uhermjakob commented 6 years ago

Annotation of earliest/earlier/early/latest/later/late is often tricky indeed. I hope that we will soon have a frameset for "late" in the AMR Editor, and maybe for "early" as well, along with additional documentation.

Meanwhile I propose the following senses for "earliest"

I have already updated all 8 AMRs that contain "earliest" accordingly. Most of those AMRs needed a lot of work. But I'm certainly open to further suggestions and improvements, particularly regarding the "earliest = first (in time)" case.

"Latest" is documented in the AMR Dict and occurs in 6 consensus/guidelines AMRs. It is annotated as (... :mod late). Recall that https://www.isi.edu/~ulf/amr/lib/amr-dict.html#last identifies two senses for "late" (1) final (as in "the last week of May" or 'Johnny Ace's final words were "It's okay! Gun's not loaded… see?"') annotated as (... :ord (ordinal-entity :value -1)) (2) most recent/previous (as in "last week", "last October" or "latest quarter") annotated as (... :mod last) "Latest" is limited to sense (2) i.e (... :mod last).

Regarding "earlier" and "later" There are already some examples in the AMR Dict along with consensus/guidelines examples, but we need additional documentation and examples, as Tim already pointed out to me yesterday.

Tim and I propose the following annotation for cases such as "earlier in 2017" or "later on Tuesday". Example:

"On August 13, 1999, Steffi Graf announced her retirement later that year."
(a / announce-01
      :ARG0 (p / person :name (n / name :op1 "Steffi" :op2 "Graf"))
      :ARG1 (r / retire-01
            :ARG0 p
            :time (d / date-entity :year 1999)
            :time (a2 / after
                  :op1 d2))
      :time (d2 / date-entity :year 1999 :month 8 :day 13))

Sometimes the :op1 of before/after will be an implied "now", sometimes it will be totally unspecified. The above is in line with precedents described under https://www.isi.edu/~ulf/amr/lib/amr-dict.html#comparative and https://www.isi.edu/~ulf/amr/lib/amr-dict.html#time

Upon further investigation, I think that there are some cases where the comparative forms "earlier" and "later" are used as downtoners (just as the superlative "earliest" is sometimes used as an intensified form of "early").

Examples (in contexts with absolutely no time reference point):

Earliest = very early; earlier = somewhat early (in some contexts).

Later = somewhat late (in some contexts). "Latest" does not quite fall into this scheme. The latest report is the most recent report, it cannot refer to the second most recent one.

uhermjakob commented 6 years ago

After some more thought and unease about annotating earliest as first in some cases, I think there is a viable more traditional approach for these cases after all. It often requires a combination of have-degree-91 and :domain with careful attention how to fill the slots.

(entity {...}
    :ARG1-of (have-degree-91 :ARG2 early :ARG3 most {:ARG5 ...})
    :domain ...)

Examples:

St. Augustine was the earliest Spanish settlement in North America.
(s / settlement
  :ARG1-of (h / have-degree-91
             :ARG2 (e / early)
             :ARG3 (m / most)
             :ARG5 (s2 / settlement
                     :mod (c / country :name (n / name :op1 "Spain"))
                     :location (c3 / continent :name (n3 / name :op1 "North" :op2 "America"))))
  :domain (c2 / city :name (n2 / name :op1 "St." :op2 "Augustine")))
There are some archaelogists who believe that two sites in Syria could be the world's earliest cities.
(b / believe-01
  :ARG0 (a / archaeologist
          :quant (s / some))
  :ARG1 (p / possible-01
          :ARG1 (c / city
                  :ARG1-of (h / have-degree-91
                             :ARG2 (e / early)
                             :ARG3 (m / most)
                             :ARG5 (c3 / city
                                     :location (w / world)))
                  :domain (s2 / site :quant 2
                            :location (c2 / country :name (n / name :op1 "Syria"))))))
The earliest to settle here was this world's largest outdoor seated bronze Buddha.
(t2 / thing
      :ARG1-of (h3 / have-degree-91
            :ARG2 (e / early)
            :ARG3 (m2 / most)
            :ARG5 (t3 / thing
                  :ARG1-of (s / settle-03
                        :ARG4 (h / here))))
      :domain (s4 / statue
            :ARG1-of (h2 / have-degree-91
                  :ARG2 (l / large)
                  :ARG3 (m / most)
                  :ARG5 (s3 / statue
                        :topic (p / person :wiki "Gautama_Buddha" :name (n2 / name :op1 "Buddha"))
                        :location (w2 / world)
                        :location (o / outdoor)
                        :consist-of (b / bronze)
                        :ARG1-of (s2 / seat-01)))
            :mod (t / this)))
nschneid commented 6 years ago

I think the "latest" as :mod last policy might be confusing because "last" is also ambiguous between "final" and "most recent". Can we express it as "most recent" instead?

uhermjakob commented 6 years ago

Yes, Nathan, "last" is ambiguous, but only the "most recent" meaning of "last" is annotated as :mod last. See https://www.isi.edu/~ulf/amr/lib/amr-dict.html#last