Discourse Connectives

[timjogorman]

Hii all! I wanted to put up a list of proposed discourse connective changes. A list of a few common connectives, with proposed senses ranked by rough frequency, is at bottom.

• Directed alternatives: I'd propose using "replace-01" and ":alternative" for connectives like "instead of", "rather than", "in place of", "in lieu of", etc. Most of these look to now be :prep-instead-of in the release set..
• Exceptions: I'd propose "exclude-01 / :exception" and used for connectives like except, besides, excluding, but for, otherwise, etc, as in

I've seen otherwise intelligent people convinced that ....

    (s / see-01
            :ARG0 (i / i)
            :ARG1 (c / convince-01
                  :ARG0 (p / person
                        :ARG1-of (i2 / intelligent-41
                              :exception c

• Disjunction: Splitting between inclusive 'or' and exclusive 'or' seems like a gargantuan task, and perhaps not always clear. However, in figuring that out, I've noticed that we tend to use an "if not X, Y" pattern for the disjunctions where that makes sense. I'd propose having some frames (particularly for "otherwise") have that suggestion -- use of have-condition-91 along with negation on the condition -- as an alternative. This is one of those things we often seem to already do:

Health and safety and environmental horrors need to be regulated otherwise industries will not EVER do whats right for humanity.

(h / have-condition-91
      :ARG1 (d / do-02 :polarity -
            :ARG0 (i / industry)
            :ARG1 (t / thing
                  :ARG1-of (r / right-42
                        :ARG2 (h2 / humanity)))
            :time (e / ever))
      :ARG2 (n / need-01
            :ARG0 (a / and
                  :op1 (h3 / horror
                        :mod (h4 / health))
                  :op2 (h5 / horror
                        :mod (s / safe-41))
                  :op3 (h6 / horror
                        :mod (e2 / environment)))
            :ARG1 (r2 / regulate-01
                  :ARG1 a)))

• I'd propose we adopt "before" as a consistent way of encoding things being in temporal sequence -- perhaps using an actual predicate like in-temporal-sequence-91 or predate-01. Part of this is the mere convenience that if you present a list of options for things like "once" or "since", one of those options should be the temporal expression.
• I'd propose using "continue-sequence-91" or ":sequential" for anything marking something in sequence that doesn't fit more meaningful categories, for things like furthermore, moreover, besides, etc.
• Finally, I'd propose that we expand exemplify-01 to all instance+generalization relationships. This is most tricky with things like

Since 9 / 11 , Middle Easterners in general , and Iraqis in particular , have faced enormous hurdles getting admitted to the United States ;

(f / face-01
  :ARG0 (a / and
          :op1 (p / person
                 :mod (w / world-region
                        :name (n3 / name
                                :op1 "Middle"
                                :op2 "East"))
                 :prep-in (g / general))
          :op2 (p2 / person
                 :mod (c2 / country
                        :name (n4 / name
                                :op1 "Iraq"))
                 :prep-in (p3 / particular)))
  :ARG1 (h / hurdle
          :mod (e2 / enormous)
          :time (g2 / get-03
                  :ARG1 (a2 / admit-02
                          :ARG1 a
                          :ARG2 (c3 / country
                                  :name (n6 / name
                                          :op1 "United"
                                          :op2 "States")))))
  :time (s / since
          :op1 (e3 / event
                 :name (n5 / name
                         :op1 "9/11"))))

Examples of sense suggestions for connectives

while: be-temporally-at.91, contrast.01, have-concession.91 however: contrast.01, have-concession.91 with respect to: concern-02 instead: replace.01, contrast.01 once: in-temporal-sequence-91, have-condition.91 since: cause.01, have-purpose-91, in-temporal-sequence-91 but: contrast.01, have-concession.91, exclude.01 otherwise: exclude.01, { have-conditional-91 :arg2 have-polarity-91 } in other words: mean.01, exemplify.01 in short: exemplify.01

[uhermjakob]

It would be great to treat "instead of", "rather than", "in place of", "in lieu of" etc. in some unified semantic way. Could you flash out your proposal with some specific examples, e.g. for

• Tim gave the book to Martha instead of to Claire.

Details (incl. exact placement) might be tricky. The sentence probably means something like "Even though one might have expected that Tim would give the book to Claire, Time gave the book to Martha." The following normal/odd sentence pair should illustrate the expectational aspect:

• Tim went for an early morning bike ride instead of sleeping in.
• Tim went for an early morning bike ride instead of drowning puppies.

For the other connectives it would also be useful to have concrete examples (as they would show up in the AMR annotation dictionary).

Thanks!

[nschneid]

I was looking something up in the Quirk et al. grammar and found that it has a nice taxonomy of adverbials, grouped under adjuncts, subjuncts, disjuncts, and conjuncts. I've uploaded some of the figures/examples in case they may be useful for the discourse discussion.

[timjogorman]

Hi all! Sorry for the delay on getting this up, and the sheer length of it. Here's a more full write-up of the whole connectives idea with more examples. Thanks for the comments Ulf, and for the pointers to Quirk et al Nathan! This is a thorny topic, so please critique, suggest and amend with abandon (especially on predicate names, I'm really horrible at that).

For those who weren't at the AMR workshop, we talked about our treatment of discourse connectives and discourse markers. I proposed that some connectives and discourse markers aren't consistently treated, and that it would be nice to have a list of senses and arguments for each connective -- like what we have for verbs. This would be particularly important in dealing with the long tail of connectives and markers that aren't very easy to define.

Here's an example frame of what I'd aim for (from "while"). These are ranked based on expected frequency, as mostly informed by the Penn Discourse Treebank. That's the second part of this idea -- that instead of hammering out frames for each connective, we use extant annotations to guide what kinds of predicates we expect a connective to link to.

While 1 - use contrast-01

Implies a contrastive relationship between two propositions, and should be symmetric (can be swapped without change in meaning) Arg1: first item in comparison Arg2: second item in comparison

Gunaratna stated a portion of the money is raised in Southeast Asia 
while some comes from the Middle East and other countries.

(s / state-01
      :ARG0 (p / person
            :name (n / name :op1 "Gunaratna"))
      :ARG1 (c / contrast-01
            :ARG1 (r / raise-01
                  :ARG1 (m / money
                        :quant (p2 / portion))
                  :location (w / world-region :wiki "Southeast_Asia"
                        :name (n2 / name :op1 "Southeast" :op2 "Asia")))
            :ARG2 (c3 / come-01
                  :ARG1 (m2 / money
                        :quant (s2 / some))
                  :ARG3 (a / and
                        :op1 (w2 / world-region :wiki "Middle_East"
                              :name (n4 / name :op1 "Middle" :op2 "East"))
                        :op2 (c4 / country
                              :mod (o / other))))))

While 2 - Temporal while.

Use :time or be-temporally-at-91.

Me? I'll cheer you on, while me and my friends get rich off your efforts."
(m / multi-sentence
      :snt1 (i / i :mode interrogative)
      :snt2 (c / cheer-04
            :ARG0 (i2 / i)
            :ARG1 (y / you)
            :time (e / enrich-01
                  :ARG1 (a / and
                        :op1 i2
                        :op2 (f / friend
                              :poss i2))
                  :ARG2 (e2 / effort
                        :poss y))))

While 3 - Concession: use :concession or have-concession-91.

The "while" clause will be the "concession" or the arg2 argument, and the main clause will be the main event that is true in spite of the concession.

Algerian Interior Minister Yazid Zerhouni stated that while Sarkozy's remarks were a sign of progress they were not sufficient.

(s / state-01
      :ARG0 (p3 / person
            :name (n / name :op1 "Yazid" :op2 "Zerhouni")
            :ARG0-of (h / have-org-role-91
                  :ARG1 (c / country :wiki "Algeria"
                        :name (n3 / name :op1 "Algeria"))
                  :ARG2 (m / minister
                        :mod (i / interior))))
      :ARG1 (s2 / suffice-01 :polarity -
            :ARG0 (r / remark-01
                  :ARG0 (p2 / person :wiki "Nicolas_Sarkozy"
                        :name (n4 / name :op1 "Sarkozy")))
            :ARG1 p3
            :concession (s3 / signal-01
                  :ARG0 r
                  :ARG1 (p / progress-01))))

While I don't think that this is absolutely necessary for "while" and other relatively easy connectives, I think that even these easy connectives would involve less cognitive load if annotators could quickly access a simple list of options. This would also provide examples of low-frequency senses that might not appear in a quick search.

This also brings up a bunch of connectives whose senses don't currently map to a current predicate, or which have one or more issues worth mentioning. I'll go through individual issues related to each kind of connective (using that loosely), presenting a proposal for each. There is a little summary at the end of proposed changes, but the general pattern is that I'm suggesting low-effort changes. The proposed predicate additions, in total, this would only involve re-examining about 150 instances (based on the release corpus), although there are probably more changes that would be entailed by adding frames, and this would potentially allow us to make further quality control checks on such connectives (with always entails more work).

The state of our sense inventory -- how we match up with PDTB

For many connectives in the corpus, our current set of predicates totally works. The sense hierarchy of PDTB is shown below, with green circles corresponding to clusters that seem completely unproblematic, such as contrast-01, :concession, :condition, and :time.

Most of the other senses that they use in PDTB either have one or two minor issues to discuss, or require addition of a predicate but one that doesn't seem that problematic:

Causation

Examples of connectives that relate to this: so, because, therefore, consequently, hence, therefore, thereby, thus, since, as

What other models and the literature say:

PDTB distinguishes between cause and pragmatic causes -- the later dealing with justifications and other points where one argument simply causes one to know, think or say the other argument, as in "[Mrs Yeargin is lying], because [they found students in an advanced class a year earlier who said she gave them similar help.]"

What we currently do:

We make the distinction between cause-01 and have-purpose-91, but also occasionally use infer-01, particularly for "so", as in:

purpose:

I need to set these things correct so that the record will be clear testimony at Rex's sanity hearing.
(n / need-01
      :ARG0 (i / i)
      :ARG1 (s / set-02
            :ARG0 i
            :ARG1 (t / thing
                  :mod (t2 / this))
            :ARG2 (c / correct))
      :purpose (t3 / testify-01
            :ARG0 (r / record)
            :manner (c2 / clear)
            :location (h / hearing
                  :mod (s2 / sanity)
                  :poss (p / person :name (n2 / name :op1 "Rex")))))

cause:

I never had dry mouth until this high dose of C, so I am assuming it is the C.
(a / assume-02
      :ARG0 (i / i)
      :ARG1 (c / C
            :domain (i2 / it))
      :ARG1-of (c2 / cause-01
            :ARG0 (h / have-03 :polarity -
                  :ARG0 i
                  :ARG1 (m / mouth
                        :ARG1-of (d / dry-02))
                  :time (e / ever)
                  :time (u / until
                        :op1 (d2 / dose
                              :mod (h2 / high)
                              :quant-of c
                              :mod (t / this))))))

infer

So it is a society with a very male testosterone-rich style.
(i2 / infer-01
      :ARG1 (s / society
            :prep-with (s2 / style
                  :mod (t / testosterone
                        :mod (m / male)
                        :mod (r / rich)
                        :degree (v / very)))
            :mod (s3 / so)
            :domain (i / it)))

I think, therefore I am .
(i / infer-01
      :ARG1 (b / be-02
            :ARG1 (i2 / i))
      :ARG2 (t / think-01
            :ARG0 i2))

What to decide:

I think that giving "purpose" and "cause" meanings to connectives is pretty easy. The question is whether to present "infer-01", or any other option, for "Pragmatic Cause" causation:

a) Do we want to allow pragmatic causes, like justification, etc.? b) If we do, should these have a special connective (such as "infer-01")?

I'd assume that we only want :cause and :purpose, but would be happy to add infer-01 if people back it.

Exceptions

Words and phrases such as: except, besides, excluding, but for, otherwise, but What we are currently doing: It looks like most of these get "except-00", with arg2 as the set and arg1 as the excluded smaller set, as in:

Iran currently administers the death penalty more than any other country besides China.
(p2 / penalize-01
      :ARG0 (c3 / country :wiki "Iran" :name (n3 / name :op1 "Iran"))
      :manner (d2 / die-01)
      :time (c2 / current)
      :mod (m2 / more
            :compared-to (c / country
                  :mod (o2 / other)
                  :ARG2-of (e / except-00
                        :ARG1 (c5 / country :wiki "China" :name (n / name :op1 "China"))))))

What to decide:

This seems to very consistently have already been annotated as either :prep-except or except-00. Except-00, furthermore, seems to have the same numbered arguments as include-91, which is very encouraging. I'd propose we just add a semantic role :exception, a reification have-exception-91 ("exclude-91" would also work if people want to link to an existing frame)

Temporal Sequences:

Examples of connectives that relate to this: once, since, when, as soon as, after, before, What we currently do: Temporal meanings for connectives like "once", "since", "before" and "after", all get coded as :time :op1 .

Traffic has more than doubled in one section of Huguenot Trail since 288 opened , reports Will Jones .
(r2 / report-01
      :ARG0 (p / person :wiki -
            :name (w / name :op1 "Will" :op2 "Jones"))
      :ARG1 (d / double-01
            :ARG1 (t / traffic
                  :location (s / section :quant 1
                        :part-of (r / road :wiki -
                              :name (h / name :op1 "Huguenot" :op2 "Trail"))))
            :time (s2 / since
                  :op1 (o / open-01
                        :ARG1 (r3 / road :wiki "Virginia_State_Route_288"
                              :name (x / name :op1 288))))
            :mod (m / more-than)))

What to decide:

I think any meaningful treatment of time should be for another day. My question is whether it's ok to stick with "time :since" and ":time once", or whether we want to convert these simple predicates into ":time before" and ":time after", or even changed to something really AMR-an like "be-in-temporal-sequence-91".

I'd propose that we don't touch them. This would involve one option for "since" be to "frame as (:time since)". We will hopefully revisit time eventually anyways, so I'd rather leave them alone then.

Adversatives

There's a class of connectives and discourse markers generally called "adversatives", which imply that some sort of implication exists -- usually from the prior discourse or the world at large -- that disagrees with what you are saying, as in "actually" in the following:

Actually, I think you are wrong on the moderate/independent voter turnout, too, but I could be wrong.

(h / have-concession-91 :ARG1 (t / think-01 :ARG0 (i / i) :ARG1 (w / wrong-41 :ARG1 (y / you) :ARG2 (t2 / turn-17 :mod (t3 / too) :topic (s2 / slash :op1 (p2 / person :mod (m2 / moderate)) :op2 (p3 / person :ARG0-of (d / depend-01 :polarity -)) :ARG0-of (v / vote-01)))) :mod (a / actual)) :ARG2 (p / possible :domain (w2 / wrong-41 :ARG1 i)))

I'd vote that these are Contrast-01 (they tend to get "contrast" in PDTB, have been related to "antithetical" in RST (which is pretty similar to our contrast-01), are generally are either contrast or counter-expectation in the contrastive/typlological ling lit) but it opens up the giant can of worms that is markers of certainty. Asserting that something is true is marked, and often implies some P implying the opposite.

What to decide:

I'm not touching epistemic marking here(though I'd totally welcome it), and would propose that connectives like "in fact" be contrast-01 only in strongly contrastive/adversative meaning, and otherwise be :mod fact.

A few messier issues

There's a few remaining parts of the PDTB inventory that have thornier questions, although most of them are managable. This is mostly around "Elaboration" and "Specification", which is a trash bin for a lot of more complicated connectives:

Chosen Alternatives ("instead of" connectives)

Words and phrases such as: instead of, instead, rather, in place of, in lieu of, etc.

PDTB defines these words as a separate class -- chosen_alternative -- which fits into a larger "alternative" class. Most of the ling. lit describes the core meaning for these much as Ulf did in his comment, as "used to deny an earlier assumption"(Thompson 1972:243) or as signally "that S1 must represent an action or a state which did not/will not occur, in contrast with S2, which represents an action or a state which did/will occur"( Frazier 2006:78).

These have two issues. The first is that there's a subset of these that actually end up having a "preference" semantics -- usually only "rather", and usually only after "would" -- which many annotators catch:

Also,to use an old cliche..."I would rather die on my feet then live on my knees"
(u / use-01
      :ARG1 (c / cliche
            :mod (o / old)
            :topic (p / prefer-01
                  :ARG0 i
                  :ARG1 (d / die-01
                        :ARG1 (i / i)
                        :prep-on (f / foot
                              :part-of i))
                  :ARG2 (l / live-01
                        :ARG0 i
                        :prep-on (k / knee
                              :part-of i))))
      :mod (a / also))

I'd propose that "prefer-01" shouldn't be anything special, but "rather" should note and allow a "prefer-01" sense. Even with "rather" this is really quite rare (Salkie 2007), so I'm not sure if we need to worry about it.

For the rest -- which is the vast majority -- we've generally been treating these using :prep-instead-of, with occasional usages of contrast-01, etc. The issue with these is what Ulf alluded to; getting the explicit semantics of what is expected to be true, and what is actually being replaced, is rather complicated. For example, ideally the following two sentences would be similar, but using :prep-instead-of (or a new, equivalent relation) they aren't:

"Tim gave the book to Martha instead of Claire"
(g / give-01
      :ARG0 (p / person :name (n / name :op1 "Tim"))
      :ARG1 (b3 / book)
      :ARG2 (p2 / person :name (n2 / name :op1 "Martha")
            :prep-instead-of (p3 / person :name (n3 / name :op1 "Claire"))))

"Tim gave the book to Martha instead of giving it to Claire"
(g / give-01
      :ARG0 (p / person :name (n / name :op1 "Tim"))
      :ARG1 (b3 / book)
      :ARG2 (p2 / person :name (n2 / name :op1 "Martha"))
      :prep-instead-of (g2 / give-01
            :ARG0 p
            :ARG1 b3
            :ARG2 (p3 / person :name (n3 / name :op1 "Claire"))))

In actual data, more of the "instead" relations are between events, and so the elidation issue is not as prominent. Suggesting that annotators copy the predicates for these kind of things isn't as clear cut as other cases, either, as I'm not sure whether (A) would be interpreted as underlyingly being (B), (C), or (D): A)"Tim decided he needed to give the book to Martha instead of Claire" B)"Tim decided he needed to give the book to Martha instead of giving it to Claire" C)"Tim decided he needed to give the book to Martha instead of needing to give it to Claire" D)"Tim decided he needed to give the book to Martha instead of deciding he needed to give it to Claire"

In the end, I'd vote to have this whole thing be a simple relation, for now (say, ":replaced-alternative" or "instead-of" with a reification (replace-alternative-91, maybe, or replace-01). The main semantics bit that we're dropping isn't any long-tail complexity, but rather that annotators mark "mod instead" instead of linking to the think being replaced.

Here's the most complicated set of "instead of" relationships I've seen so far:

Upon his return to Beijing, the situation in Urumqi was far from settling, instead the *Uyghur terrorists* becoming even more rampant, and the numbers of attacks and bombings were increasing rather than decreasing!
current:
    :snt2 (c2 / contrast-01
            :ARG1 (s / settle-03 :polarity -
                  :ARG1 (s2 / situation
                        :location (c3 / city
                              :name (n3 / name :op1 "Urumqi")))
                  :degree (v2 / very)
                  :time (r / return-01
                        :ARG1 (h2 / he)
                        :ARG4 (c4 / city :wiki "Beijing"
                              :name (n4 / name :op1 "Beijing"))))
            :ARG2 (a / and
                  :op1 (b / become-01
                        :ARG1 (t3 / terrorist
                              :mod (e3 / ethnic-group
                                    :name (n5 / name :op1 "Uyghur")))
                        :ARG2 (r2 / rampant
                              :degree (m3 / more
                                    :mod (e2 / even))))
                  :op2 (i3 / increase-01
                        :ARG1 (n6 / number
                              :quant-of (a2 / and
                                    :op1 (a3 / attack-01)
                                    :op2 (b2 / bomb-01)))
                        :prep-instead-of (d2 / decrease-01
                              :ARG1 n6))))
proposed:
(a4 / and
      :op1 (h / have-polarity-91
            :ARG1 (s3 / settle-03
                  :ARG1 (s2 / situation
                        :location (c3 / city :name (n3 / name :op1 "Urumqi")))
                  :degree (v / very)
                  :time (r / return-01
                        :ARG1 (h2 / he)
                        :ARG4 (c4 / city :wiki "Beijing" :name (n4 / name :op1 "Beijing"))))
            :ARG2 -)
      :op2 (r5 / replace-alternative-91
            :ARG1 (i3 / increase-01
                  :ARG1 (n6 / number
                        :quant-of (a2 / and
                              :op1 (a3 / attack-01)
                              :op2 (b2 / bomb-01)))
                  :replaced-alternative (d2 / decrease-01
                        :ARG1 n6))
            :ARG2 s3))

I'd thus propose that most of our issues with this can be treated simply with a predicate: replace-alternative-91 :arg1 new argument :arg2 replaced argument, with X :replaced-alternative Y being X :arg1-of replace-alternative-91, :arg2 Y

What to decide:

The main question is, does anyone have a better representation of these, or wish we dealt more deeply with the semantics? For phrases such as the sentence above "I gave the book to Martha instead of Claire", is there a desire to annotators to add a copy (i.e. "instead of giving it to Claire" )?

Some test sentences for anyone who wants to build a better frame for this: The sky was red instead of blue I didn't get cake but instead got pie

"Otherwise"-style Disjunctions

Words and phrases such as: Otherwise, lest, unless, if not What we are currently doing: These are often represented, if not using "mod otherwise", using have-condition-91, with polarity -, as in:

The game will be canceled unless the rain stops .

(c / cancel-01 :ARG1 (g / game) :condition (s / stop-01 :polarity - :ARG1 (r / rain-01)))

The exception being this one treatment: Of course, don't go over the top with it lest you go in completely the opposite direction. (g / go-01 :polarity - :ARG2 (t / top :ARG1 (y / you) :mod (o / over)) :prep-with (i / it) :mod (o3 / of-course) :purpose (g2 / go-01 :polarity - :ARG1 y :direction (d / direction :ARG1-of (o2 / opposite-41 :degree (c / complete)))))

What to decide:

This is something that annotators seem to clearly need guidance on. Adding a symmetric exclusive "or" doesn't seem like it would capture the implied semantics, either. Since annotators have already done "have-condition-91" for this, and it seems like a good treatment, I propose that that simply be included in a frame.

Exemplify, Instantiation and Summarization

PDTB "Instantiation" is a clear set that maps to exemplify-01 in AMR, as in: for example, for instance, such as, e.g. What we currently do: Naturally we use exemplify-01 for these. More specifically, I would note that many instances of exemplify-01 are generalizations with instances or arguments (rather than being of the "fruit, for example bananas", set/subset variety), as in "for example" below:

Indians in particular are treated badly, there are Indian maids and they are often hit by the house wife if they for example make a small mistake or sometimes they are hit with no reason.

(e2 / exemplify-01
      :ARG0 (o3 / or
            :op1 (h / hit-01
                  :ARG0 (w / wife
                        :mod (h2 / house))
                  :ARG1 (m / maid
                        :mod c)
                  :frequency (o / often)
                  :condition (m2 / mistake-01
                        :ARG0 m
                        :ARG1 (t2 / thing
                              :mod (s / small))
                        :ARG0-of (e / exemplify-01)))
            :op2 (h3 / hit-01
                  :ARG1 m
                  :frequency (s2 / sometimes)
                  :prep-with (r / reason :polarity -)))
      :ARG1 (t / treat-01
            :ARG1 (p / person
                  :mod (c / country :name (n / name :op1 "India")))
            :manner (b / bad)
            :manner (p2 / particular)))

The fact that these get used for generalization/instantiation predicate raises the question about what to do with "summarization/generalization" words like 'in other words', 'in short', 'in summary', 'in the end', etc. What I wonder is whether we can make them all exemplify-01 as well.

Currently, we tend to treat "in other words" as mean-01 when it's clearly restatement, treat "in summary" as summary-01, and punt on everything else. As example of the summarization usage (googled for context):

"I suspect you are wrong on the voting inclinations of the moderates, and especially the ones you call tea baggers. Actually, I think you are wrong on the moderate/independent voter turnout, too, but I could be wrong. I'm expecting a fairly good anti Obama turnout amongst that group.

I could accept Hillary. Like Phoenix, I'm not greatly impressed by the Republican field. Not crazy about the presumed Democratic candidate either. On the other hand, Obama stepping aside seems even less likely than Cain giving up.

I guess that in summary, I'm not optimistic about the future of the country for the next five years."

(g / guess-01
  :ARG0 (i / i)
  :ARG1 (o / optimistic
          :polarity -
          :domain i
          :topic (f / future
                   :poss (c / country)
                   :duration (t2 / temporal-quantity
                               :quant 5
                               :unit (y2 / year
                                       :mod (n / next))))
          :ARG1-of (s / summarize-01)))
'''
Another very important issue is with unemployment so high, how can we strengthen the Fed's full-employment mandate and ensure that it conducts monetary policy to achieve maximum employment? In other words the Fed has a number of mandates and one of them is to control inflation. But one of them also is to pursue policies that lead to full employment. Is the Fed doing that in an adequate way? Well I would argue that by definition when you have 16% of people unemployed and underemployed it really is not.

In other words the Fed has a number of mandates and one of them is to control inflation.

(a / and :op1 (h / have-03 :ARG0 (g / government-organization :wiki "Federal_Reserve_System" :name (n / name :op1 "Federal" :op2 "Reserve")) :ARG1 (m / mandate :quant (n2 / number))) :op2 (c / control-01 :ARG0 g :ARG1 (i / inflation) :ARG1-of (i2 / include-91 :ARG2 m :ARG3 (o / one))) :mod (w / word :mod (o2 / other))) '''

Merging those generalization connectives into exemplify-01 would let us get roughy similar treatments for sentence pairs like the folowing:

"I was having a horrible day; for example, I lost my keys and then missed the bus" "I had lost my keys, missed the bus, my banana for lunch was all brown; in sum, I was having a horrible day"

What to decide:

Does anyone have opinions on this, such as prefering a summarize-01 frame?

The long tail of conjunctions, topic manipulation and related markers:

Additives: Furthermore, in addition, moreover, additionally, further Topic shifting/return: so, now, yeah, by-the-way, Topic dismissal/c: anyways, anyway, clause-final still Others: similarly, likewise

What we do with these now: "and" is clearly "and". Some of the "additives" are treated as "and", and everything else is done with :mod coding, or in unfortunate references to other meanings of those connectives (topic-shifting "so" is usually "cause-01", etc.).

Furthermore, the many strategic adjustments and ideas of imperialist America are all aimed at the rising China.

(a5 / and
      :op2 (a / aim-02
            :ARG1 (a2 / and
                  :op1 (a3 / adjust-01
                        :ARG0 (c2 / country :wiki "United_States"
                              :name (n2 / name :op1 "America")
                              :mod (i2 / imperialism))
                        :ARG1 (s / strategy))
                  :op2 (i / idea
                        :poss c2)
                  :quant (m / many)
                  :poss c2)
            :ARG2 (c / country :wiki "China"
                  :name (n / name :op1 "China")
                  :ARG1-of (r / rise-01))
            :quant (a4 / all)))

I imagine that we will end up running into many infrequent terms used for discourse usages like this. Making clear distinctions with these terms is probably beyond us, and doesn't seem to be settled in the literature for most of these, either.

What to decide:

Option 1: Drop them all. Have suggestions in sense lists, with examples, of what kinds of usages for a given connective should be dropped.

Option 2: Have a catch-all discourse "bin" for these, so that annotators don't spend time trying to figure out how to do them. I would propose having a three-place predicate, with a place for the actual connective, as in: have-sequential-relationship-91: For any sort of rhetorical, narrative or conversational component. arg1: prior item in an implied sequence, or proposition being continued, shifted away from, etc. arg2: secondary item in a sequence arg3: type of shift (concatenate words with hyphens)

Clinton had a women secretary of state, and Rice does not represent blacks, so what is your beef?

currently: (a / and :op1 (h / have-03 :ARG0 (p / person :name (n / name :op1 "Clinton")) :ARG1 (w / woman :ARG0-of (h2 / have-org-role-91 :ARG2 (s / secretary :mod (s2 / state))))) :op2 (r / represent-01 :polarity - :ARG0 (p2 / person :name (n2 / name :op1 "Rice")) :ARG1 (p3 / person :mod (b / black))) :ARG2-of (i / infer-01 :ARG1 (b2 / beef-00 :ARG0 (y / you) :ARG1 (a2 / amr-unknown))))

proposed: (a / and :op1 (h / have-03 :ARG0 (p / person :name (n / name :op1 "Clinton")) :ARG1 (w / woman :ARG0-of (h2 / have-org-role-91 :ARG2 (s / secretary :mod (s2 / state))))) :op2 (r / represent-01 :polarity - :ARG0 (p2 / person :name (n2 / name :op1 "Rice")) :ARG1 (p3 / person :mod (b / black))) :ARG1-of (c / have-sequential-relationship-91 :ARG2 (b2 / beef-00 :ARG0 (y / you) :ARG1 (a2 / amr-unknown)) :ARG3 (s3 / so)))

(BTW, he's very thin, but has a heart condition).

currently: (t / thin :domain (h / he) :concession-of (h2 / have-03 :ARG0 h :ARG1 (c / condition :mod (h3 / heart))) :degree (v / very) :mod (b / by-the-way))

proposed: (t2 / thin :domain (h4 / he) :concession-of (h2 / have-03 :ARG0 (h / he) :ARG1 (c / condition :mod (h3 / heart))) :degree (v / very) :ARG2-of (h5 / have-sequential-relationship-91 :ARG3 (b / by-the-way)))

Obviously this doesn't encode what's going on in a deep way, but it might be pretty flexible and allow us to capture a lot of otherwise wince-inducing treatments of discourse phrases:

It might , you see , have been a new kind of baobab .

current: (p / possible :domain (b2 / baobab :mod (k / kind :mod (n / new)) :domain (i / it)))

possible: (p / possible :domain (b2 / baobab :mod (k / kind :mod (n / new)) :domain (i / it)) :ARG2-of (h / have-sequential-relationship-91 :ARG3 (s / see-03 :ARG0 (y / you))))

On What I Propose to Not Treat at all :

I propose completely punting on, and therefore leaving as-is, various kinds of "focus marking", including connectives, such as "in particular", "let alone", and discourse markers like "even", as in: "John likes fruit, and bananas in particular" "John likes fruit in general, and especially likes banana" "John doesn't even like apples, let alone jicama" "John wouldn't even try apples much less eat your fruit salad"

These all relate to these "scalar models" systems, as first proposed in this paper by Fillmore on the "let alone"-type connectives (Fillmore 1988). Even more than the sticky issues with "instead of", the semantics of these get really hairy, and I think we should go back to them if anyone gets a good solution.

Some Example Frames:

Though 1 - Use have-concession-91;

use this both for subordination (when the "though" clause is might be assumed to negate or prevent the main event, but didn't) and for clause-final though, which usually has an implied concession.

If you're confused (though I can't see why you would be) this is what I am talking about

(t / talk-01
      :ARG0 (i / i)
      :ARG1 (t2 / this)
      :ARG1-of (h / have-condition-91
            :ARG2 (c / confuse-01
                  :ARG1 (y / you))
            :ARG1-of (h2 / have-concession-91
                  :ARG2 (p / possible
                        :domain (s / see-01
                              :ARG0 i
                              :ARG1 (c2 / cause-01
                                    :ARG1 (c3 / confuse-01
                                          :ARG1 y)))
                        :polarity -))))

Did n't make much sense , though .

(h / have-concession-91
      :ARG1 (m / make-01
            :ARG1 (s / sense
                  :degree (m2 / much :polarity -))))

Though 2 - Use contrast-01

connects two elements that are asserted to normally be incompatible: ARG0: entity making comparison agent ARG1: first item in comparison patient1 ARG2: second item in comparison, usually prepositional patient2

I'm no expert, but Nadia Comenechi, the famous Romanian gymnast, has no health problems from her gymnastics days though her former gymnast husband is wracked with arthritis.

(c / contrast-01
      :ARG1 (h / have-03
            :ARG0 (g / gymnast :name (n / name :op1 "Nadia" :op2 "Comaneci")
                  :mod (c2 / country :wiki "Romania" :name (n2 / name :op1 "Romania"))
                  :mod (f / famous))
            :ARG1 (p / problem
                  :mod (h2 / health)
                  :source (d / day
                        :mod (g2 / gymnastics)
                        :poss g))
            :ARG1-of (h3 / have-concession-91
                  :ARG2 (e / expert-41 :polarity -
                        :ARG1 (i / i))))
      :ARG2 (h4 / husband
            :time (f2 / former)
            :poss g
            :ARG1-of (w / wrack-01
                  :ARG0 (a / arthritis))))

still 1 - use have-concession-91

use when "still" entails that its argument (arg1) exists despite some other facts implying that it would not. arg1: main event arg2: concession

and I 'd still pay to live in the country .
(a / and
      :op2 (p / pay-01
            :ARG0 (i / i)
            :ARG3 (l / live-01
                  :ARG0 i
                  :location (c / country))
            :ARG1-of (h / have-concession-91)))

still 2 - temporal contrast "still"; use :time still.

This is for the specific sense implying that the time of the event contrasts with some expectation that the event should have concluded by that point.

Somehow it still feels heavy on my face.

(f / feel-05
      :ARG1 (i2 / it)
      :ARG2 (h / heavy)
      :manner (s / somehow)
      :location (f2 / face
            :part-of (i / i))
      :time (s2 / still))

yay for the French...at least one western democracy opposes communist police states still.

(c / cause-01
      :ARG0 (o / oppose-01
            :ARG0 (d / democracy :quant 1
                  :mod (l / least)
                  :location (w / world-region :name (n2 / name :op1 "West")))
            :ARG1 (s2 / state
                  :mod (p / police
                        :mod (c2 / communism)))
            :time (s / still))
      :ARG1 (y / yay
            :prep-for (p2 / person
                  :mod (c3 / country :name (n / name :op1 "France")))))

still 3 - use contrast-01.

Use if "still" seems to contrast two predications that could be incompatible with each other. ARG0: entity making comparison agent ARG1: first item in comparison patient1 ARG2: second item in comparison, usually prepositional patient2

I stop the multivitamin and my mouth symptoms are gone, sadly my eyesight is still damaged.
(c / contrast-01
      :ARG1 (a / and
            :op1 (s / stop-01
                  :ARG0 (i / i)
                  :ARG1 (m2 / multivitamin))
            :op2 (g / go-01
                  :ARG1 (s2 / symptom
                        :mod (m3 / mouth
                              :part-of i))))
      :ARG2 (e / eyesight
            :ARG1-of (d / damage-01
                  :mod (s4 / sad))))

One Indian government official who studied the blueprints stated the science was old and dated but still functional.
(s / state-01
      :ARG0 (p / person :quant 1
            :ARG0-of (h / have-org-role-91
                  :ARG1 (g / government-organization
                        :ARG0-of (g2 / govern-01
                              :ARG1 (c / country :wiki "India"
                                    :name (n / name :op1 "India"))))
                  :ARG2 (o3 / official))
            :ARG0-of (s2 / study-01
                  :ARG1 (b / blueprint)))
      :ARG1 (s3 / science
            :mod (c2 / contrast-01
                  :ARG1 (a / and
                        :op1 (o2 / old)
                        :op2 (d / dated))
                  :ARG2 (f / function-01
                        :ARG0 s3)))) 

in turn, sense 1: have-sequential-relationship-91

Mark this as a narrative sequence relationship whenever possible.

The North Korean government in turn claimed to be in desperate need of power generation after the oil supplies were cut and began during 021215-021221 to unseal the spent fuel rods at the facility and to dismantle monitoring equipment.

(a2 / and
      :op1 (c / claim-01
            :ARG0 (g / government-organization
                  :poss (c2 / country :wiki "North_Korea"
                        :name (n / name :op1 "North" :op2 "Korea")))
            :ARG1 (n2 / need-01
                  :ARG0 g
                  :ARG1 (g2 / generate-01
                        :ARG1 (p / power))
                  :manner (d / desperate))
        :ARG2-of (h / have-sequential-relationship-91
        :ARG3 (i / in-turn))
            :time (a / after
                  :op1 (c3 / cut-01
                        :ARG1 (s / supply-01
                              :ARG1 (o / oil)))))
      :op2 (b / begin-01
            :ARG0 g
            :ARG1 (a3 / and
                  :op1 (u / unseal-01
                        :ARG1 (r / rod
                              :consist-of (f / fuel)
                              :ARG1-of (s2 / spend-00))
                        :location (f2 / facility))
                  :op2 (d5 / dismantle-01
                        :ARG1 (e / equipment
                              :ARG0-of (m / monitor-01))))
            :time (d2 / during
                  :op1 (d6 / date-interval
                        :op1 (d3 / date-entity :year 2002 :month 12 :day 15)
                        :op2 (d4 / date-entity :year 2002 :month 12 :day 21)))))

in turn, sense 2: temporal reading

If "in turn" seems to mainly have a temporal reading, mark "in turn" using ":time after :op1 ", with an index to the prior proposition"

Example (none in AMR yet): " He was quickly on the phone with Mr. Mullins , who in turn was talking with the chairmen of the New York and Chicago exchanges . "

Finally sense 1: temporal

temporal sense referring to an undefined end point of something in the discourse, or referring to something happening in contrast to an expected end time. Use ":time final"

What do we have to dry up and die before we finally figure out what ails us?

(c / cause-01
      :ARG0 (a3 / amr-unknown)
      :ARG1 (o / obligate-01 :mode interrogative
            :ARG2 (a / and
                  :op1 (d / dry-02
                        :ARG1 (w / we))
                  :op2 (d2 / die-01
                        :ARG1 w)
                  :time (b / before
                        :op1 (f / figure-05
                              :ARG0 w
                              :ARG1 (t / thing
                                    :ARG0-of (a2 / ail-01
                                          :ARG1 w))
                              :time (f2 / final))))))

Finally sense 2: Last point in a list of arguments If "finally" marks the last point in a list of points or arguments, mark that point as :li -1

Finally, the contradictions are bound to intensify, making the situation out of hand.
(i / intensify-01
      :ARG1 (c / contradiction)
      :ARG0-of (m / make-02
            :ARG1 (c2 / control-01 :polarity -
                  :ARG1 (s / situation)))
      :li -1
      :mod (b / bound))

Summary of proposed changes and re-annotation effort:

• I'm proposing that each connective get a list of recommended treatments.
  effort: unknown; Whatever it takes Ulf to make these into amr frames (or whatever other presentation format we want)
• Replace "rather", "instead of", etc. with have-chosen-alternative-91 / :replaced-alternative; "rather" can sometimes be "prefer-01".
  effort: moderate; "prep-instead-of" could be dealt with automatically. There are 8 ":mod instead" instances in the release that one might check manually, as they are probably missing an argument, and another 50 of "rather".
• "Otherwise"-style Disjunction: I'm proposing that we use an 'if not X, Y', treatment for these. We seem to already do this already, often.
  effort: low; There have 33 instances total for "otherwise", "besides" etc in the release search.
• Exceptions: Propose mapping to :exception; I think there are 11 of these that don't already have :prep-except or exception-00.
• Summarization ("in short", "in sum", etc); I propose lumping summarization into exemplify-01 (usually as :arg1-of exemplify-01). effort: low; less that 15 of these total.
• Conjunctions and hard-to-describe topic shifting markers ("furthermore", discourse "so", "anyway(s)", etc): I propose putting them in a catch-all predicate have-sequential-relationship-91, so that annotators don't try to figure out what to do with them, whether to drop them, etc. Effort: moderate: I see 52 instances to examine (43 additives; futhermore, moreover, additionally and in addition, and 9 from topic-shifting "so"), which seems doable.

[nschneid]

I was looking at the treatment of still in the data, and rediscovered this issue.

The treatment is not consistent, and I agree something more semantic is desirable. Right now it is often :mod, but sometimes :time and even (in 2 release AMRs) the ARG2 of contrast-01!

@timjogorman proposes 3 senses, mapped to have-concession-91, :time, and contrast-01. I'm not sure I see the difference between the first and third of these.

There are cases which I think are primarily temporal, but further suggest a contrast ("I stop the multivitamin and my mouth symptoms are gone, sadly my eyesight is still damaged."). There are also other cases which are clearly not temporal—e.g., the headline Yes, I’m a nine-year-old girl. But I’m still a serious reporter.

Could we test for :time with an "it remains the case that..." paraphrase?

I think there are also cases that are similar to the temporal sense, but not quite temporal. Imagine the discourse: "Landlord: I can give you a discounted rate of $2000/mo. Prospective tenant: No, $2000/mo is still too expensive for me." Maybe this is discourse-time? Or something about scales?

[timjogorman]

I'd be fine with any treatment of "still" that gets us out of doing ":mod (s / still)"! My assumption of splitting "still" into concession and contrast-01 was largely data-driven; PDTB annotations seem to split it roughly down the middle.

As for the gradient of differences between temporal and concessive still: I'd love for us have (one day) a rich treatment of these "phasal particle" or "temporal expectation" operators, like still, already, yet/not yet, no longer, anymore (and deja, schon, noch, etc). But unless someone has some brilliant proposal for what that would be, a test like "it remains the case that..." sounds like the only easy boundary to draw.

"Or something about scales?"

See Michaelis '93 for a scalar-model treatment of still :). It has some "scalar" examples that seem similar to that:

• "I can still beat Paul. Peter is too good for me."
• "Compacts are still pretty safe; subcompacts start to get dangerous"