Gerunds (aka Closed Clausal Complements, cont)
Ann: Them/their voting early surprised me. Is that them/their ... is that the real their or not. Does anyone know the literature?
Ann: To me it looks like some sort of residual case marking mistake that English has just got as opposed to a genuine possessive.
oe: 108 & 109 are potentially different in that regard.
Ann: There are some examples where you don't seem to be able to do the their.
Emily: Rob Malouf's thesis--nominal and verbal gerunds. If you add certain things, then it's definitely nominal/verbal and that constrains the case on the specifier.
Dan/oe: It's the nominal one that requires possessive/disallows accusative.
Ann: Just wondering if you want the possessive relation in that. My feeling is not. Is that case marking that doesn't correspond to a meaning difference.
Silvie: In Quirk's grammar, it's just treated as a funny little thing. Gerund clause with subject in either case.
oe: Some of you grouped these into paraphrase sets while annotating. But I don't immediately recall the grouping. Who gave the same to all three?
Johan: The last two, maybe even all three.
Dick: Not the same, but very closely related.
Alex: Different how?
Dick: The one's with the possessive I've got the argument substitution with a residual possessive.
Alex: With them voting you don't fill the argument position?
Dick: There I do it too. But the residual possessive doesn't show up.
Dan: Same with me --- I don't have an extra 's, extra set of possessive pronouns.
oe: I think we did say that we might want that actually.
Dan: Then one has to be very careful to make it exactly right---all idiosyncrasies of the various pronouns e.g. singular their.
Emily: Can we find examples where we want to underspecify the relationship between the possessive pronoun and the gerund.
Ann: I think if you put adverb early in I think you can't.
Dan: Can we get a by phrase in there? The voting early by the students... Any voting early by students is going to be problematic. That suggests that it is possible to put the by phrase to fill that semantic argument position.
oe: If it were the case that you can't have the possessive determiner when you have the by phrase, isn't that an argument for treating the possessive determiner as unrelated to possession.
Dan: If one could demonstrate that that is really impossible, but one counterexample won't do it for me. Their voting early by the student...
Ann: Can you remember the Quirk argument for this, or do they just state it?
Silvie: I think they just gave it.
Ann: In possessives in compounds, they give strong arguments for it not being possessive. So you need something that possessive in the grammar for those examples.
Dan: Like German s in compounding, yes.
Emily: In that case you don't get the pronouns.
Ann: If you paraphrase this in German, that's genitive (not that this is a strong argument)?
oe: You are probably a better judge of German than I am.
Alex: Their getting more fit by the their personal trainer. not good even with the causative reading.
Dick: Non-gerund nominalizations: Carthage's destruction by Rome./*Carthage's destroying by Rome.
Dan: Presumably if there's a by marker, it's because of a passive form. No reason to expect the by marking more generally. And there is no ing form of the passive.
Dick: the voting by the students
Ann: Do you want that to be passive? There's some generalization that I don't remember that when you get the possessive and the by in the verbal gerund, it's always supposed to go... Cleo might know the right references.
Alex: To me as a consumer of the grammar, having the possessive here could screw up discourse processing, since it's a marker that's relevant for presuppositions.
Dan: This week's voting early by the students was unexpected. There's a reason to have the possessive without interfering with the arg-st of the verb.
oe: Don't you always promote a verbal gerund eventually to something nominal?
Dan: Only in the very last step.
Alex: What does that example show?
Dan: It shows that the possession is different from the argument filling. It's a hint that there might have to be possession as another possible reading. Maybe another ambiguity.
oe: I suspect we'll go back to that generalization because the possessive can also express the ARG2 in some cases: The symphony's performance by the orchestra.
Ann: What's confusing me is I can't remember the Grimshaw data as opposed to my potential counterexamples to the Grimshaw data.
Emily: Back to the paraphrase set question:
oe: Johan groups 6&7 and 8&9.
Woodley: I would have said that issues of aspect aside, all four are paraphrases.
[Dan teaches oe something about how to use emacs.]
Johan: This is what I did last week... vote(e2) v. voting(e2) but in the background knowledge it would be the same thing. Boxer get this from the stemmer.
oe: Should I change it?
Johan: Using of(e2,x2) from possessive, but it can resolve to any role. If I have the cake surprised me the cake isn't propositional, but maybe you can coerce.
Alex: p1 (the voting) should be presupposed in all four cases. If you negate the sentence, there's still voting.
Ann: There's two readings to this sentence. Which one are we looking at? Their early voting surprised me. The fact that they voted early v. the outcome of their voting.
Alex: There is that ambiguity, but what's potentially presupposed is that they voted early, not the result. I think in either reading. Don't you?
Ann: No, not necessarily.
Alex: Even if it's the result that surprised you---and you negate the sentence---that they voted early projects out.
Dick: I think that aspect might be important. Their vote surprised me. v. Their early voting surprised me.
Dan: Their voting early surprised me---you can't talk about the result. So there's an asymmetry there.
Emily: That's interesting...
Alex: Still presuppose that voting happened, even if you're talking about the result. So I'd put the p1 on the left of the +.
Johan: I don't disagree. There's also nested presuppositions, because possessive is also a presupposition.
Alex: You get it with Them voting early too --- it's surprise here that's the trigger. Even if their is possessive (and that's in dispute), p1 still goes up.
Dan: Maybe the hang-up is in trying to assign too specific an assumption to the of from 's. We need that to be underspecified in ordinary NPs too: Elvis Presley's movies were better than mine.
Alex: That's why I was trying to do this causative where the their could fill another role in the structure.
Dan: That's what I thought this week's would let me do.
Emily: I'm wondering if this helps with the presupposition problem though. If all possessives are actually underspecified in the role they express, they still have presuppositions. Maybe that comes from definiteness?
Dan: We could maybe treat the definiteness as a separable property.
oe: To sum up, if we take the boxes given at face value, the only difference is the labeling of the of role.
Johan: But if the whole thing is presuppositional, then there's nothing in the p1 box...
Alex: What's wrong with that?
Johan: Two triggers in the lexicon that both trigger a presupposition.
Alex: If you were doing it in Rob van der Sand's style:
p1: |e2 | |..... | |vote(e2) | |early(e2) | | ________ | | x3 || |.... || |of(e2,x2)|| ------------
oe: Suggests presupposition subcommittee, to meet at the coffee break.
Tim: All four examples treated the same in AMR, because AMR loses a lot of aspect type stuff. be-temporally-at-91 = `the earliness of their voting surprised me'. The other AMR treatment is treating voting the proposition as the thing surprising. Since we don't deal with gerunds and do possession as whatever semantic role shows up, their is always the agent.
Emily: What's the thing about aspect --- I don't see that as salient in these examples.
Woodley: I think I just made it up.
Tim: Certain relations like time can be turned into nodes with variables so that you can reference them. Any relation can be referenced if the annotator feels the need.
oe: This is what you call the reification of this relation--this role? e2 / early is a point in time?
Dan: It's a little sharper with a day like today or tomorrow.
Alex: With them voting early, e / early becomes an AR2.
oe: You continue these two equivalent?
Tim: Yeah. be-temporally-at-91 and :time early are equivalent. It's just changing the thing that's linked to surprise.
oe: Do you continue them interchangeable?
Tim: Reification v. relation yes, those are interchangeable.
Woodley: For that embedded section, it's equivalent, but it exposes a different piece of it for the surprise predicate to grab ahold of.
Johan: It's a bit like a FrameNet thing.
oe: The earliness is the state of being early, or being located at an early time.
Tim: Cut and pasted for all the others.
Emily: All four of them?
Dan: There's two representations, but that ambiguity is present for all four.
oe: Silvie, what grouping falls out of your annotations?
Silvie: We make the same trees for all four, drawing on Quirk who said that the possessive has no relevance there. Even vote would be the same. Annotation error: early should be under voting everywhere. And the lemmas should be the same everywhere. Here we lose a piece of information that this they was realized as a possessive.
Dan: Which might just be a morphological change.
Alex: Is anyone capturing the ambiguity that Ann pointed out about the result v. process in the versions that have it and those that don't. That's something like you'd like to capture in a pipeline where the discourse processing doesn't need to look at word order.
oe: I still struggle to get the result reading with voting.
Ann: There's some sort of results coming in for an election, and we're looking at the results from some constituency, where the early voting is coming back conservative in a labor constituency. Their early voting surprised me.
oe: Thanks that helped.
Alex: But you can't get at that reading with Their voting early surprised me.
Woodley: So the result reading comes from a nominal gerund?
Ann: The normal story is that the result is genuinely a nominal, as much of a noun that a gerund can ever get. The resultative is always supposed to be a real noun---can be an event denoting thing in some circumstances. Can quantify it much of the early voting surprised me.
oe: I even have the impression of an intonational contrast.
Ann: I thought you put early in to make sure we only got the verbal gerund reading.
Dan: Yes, early is convenient because it's the same shape for adj & adv, and it's a post-verbal adv.
Alex: But no one's capturing that.
Dan: No, but it's possible that the machinery is there because we follow Malouf and make the nominal gerunds into nouns lexically and the verbal gerunds only at the last step.
oe: Looking at the ERSs, we see that early is decomposed.
Dan: As in AMR.
oe: We'll see that in...
Dan: All four of them.
oe: Including the adjective?
Ann: No, I don't think the adjective.
Dan: Right the adjective is just early_a.
Emily: Why is early the adjective decomposed but not early the adverb?
Dan: It doesn't really matter here. I want early to consistently modify things. In the adjective use, there is an entity/thing. With the adverbial use, the thing has to be the time.
[Discussion of which view on MRSs to look at.]
Dan: I don't have that much new to say.
oe: You nominalize in all four, but there should be a context where we get to the possessive forms.
Dan: Yeah, there's that poss_rel that we could interpret, Parson's-style, as the most underspecified role.
oe: Here  you have actually bound the ARG1, so doing like what Dick's doing.
Dan: Right, that's suboptimal. So I'm not sure why I left the poss in.
Ann: I think it's probably better to leave it underspecified.
Dan: Right---I should not bind the ARG1, and then I can explain why the poss_rel is there.
Alex: So then why is Their loving by Mary is crap?
Ann: What do you think about The tree's deliberate felling by the farmers?
Alex/Dan: That's nice.
Ann: That was my counter-example to Grimshaw.
Alex: Right, so the possessive should be there and the binding not.
Woodley: I'm just slightly skeptical of felling.
Ann: The reason the felling example is good is to do with aspect. It's something that can happen over and over again. The reason I put deliberate in is to force the verbal gerund.
Dan: But that's an adjective. It would have to be The tree's felling deliberately.
Ann: So it's not really the nominal/verbal gerund thing.
oe: I think the subtext here is to try and work out whether the grammar forces the role assignment there with the possessive/funny case. Seems inconclusive to me from examples so far.
Alex: I'm really struggling to get the possessive to fill the object role (except in the felling case), and if this is right my intuition is that it would be easier to do that. Why is it so salient to make the possessive the agent?
oe: As a kid I climbed these trees every day. Their felling by the city authorities broke my heart.
Woodley: There may be something distinct about that category of verb.
Ann: As I remember that was a counterexample to a claim in the literature. So there may be something funny about that.
Dan: Assume that's an English sentence, and we think that it's a gerund form of the verb fell, then that possessive marked phrase must at least admit the patient role. And then it seems even more tempting, alongside This week's early voting by the conservatives it suggests that possessive marker must be prepared to assign a vague role.
Alex: I agree with all of that, but star to Their early voting by Mary.
Dan: But completeness/coherence. You have to give a different role for that to fill.
Dan: Maybe we don't have to solve this problem, but rather can say that we just don't want to decide too early. The possessive version says there's a role here that's got to be filled.
oe: Doesn't that decide which side you take in the Grimshaw/Copestake debate here?
Dan: I think Grimshaw can't deny the grammaticality of that sentence.
Ann: I think she did actually...
oe: Or come up with another property of that class of verbs, or...
Ann: If you get this in some sort of resultative reading and it's an ordinary noun then you are supposed to get all those real things.
Ann: Even if it's a noun form you can still block the result reading with deliberate, that's what I was trying to do with deliberate there.
Dan: I throw the nominalization in that others don't do, which
oe: These rules are really important to us. We all must follow them. Their repeated breaking troubles me.
Dan: Their repeated breaking by the children troubles me.
Woodley: Star. The vase broke/they broke the vase, *The rule broke/They broke the rule. And star The rule broke by the children.
Dan: Right, but we've noticed that gerunds can have something like a passive reading even without the passive morphology. Stephan's looking for one more example beyond fell.
Ann: That's the right kind of example to be looking at. Rules, like trees, are so clearly non-agentive.
oe: Now I can let go of that. Back to 109. Here the adjective early is not decomposed. Is that the only difference to the preceding one? Here you don't bind the ARG1.
Dan: It would be nice to see if there's a way to tease out that result/process ambiguity.
oe/Emily: So you could decompose early here (though you don't have to). The voting (noun) could be at an early time just like the vote_v is an early time.
Dan: That's not very relevant.
Ann: Actually, here it might be. There's something go on about the results v. the voting being at an early time.
Dick: There's not much to say that hasn't been said. The only difference here is the veridicality judgments, and it does take note of the progressive aspect but maybe it shouldn't.
Emily: So the veridicality is the surprise presupposition.
Alex: So no possessive?
Dick: Not in them voting, but in their voting.
Emily: So like we said before, you like Dan are putting the possessive in and binding the argument.
Emily: Can we revisit the underspecification question then? If the solution here is the underspecified role name.
Dick: `of' isn't underspecified, it's just non-specific.
Dan: Can you show us your semanticist card again? You'll need to have that validated.
Alex: I think there is a difference between underspecified and non-specific.
Dick: Thanks, Alex.
Ann: As I said yesterday, underspecified with respect to what? If you have a formalism that just doesn't want to say anything about it, it's non-specific. MRSs are non-specific, they're just underspecified with respect to certain kinds of logical formulas.
Alex: Or anything else.
Ann: But I think I want to decide that the target language is a non-specific one.
Johan: What about the intersective modifier?
Dick: AMOD relation that applies to both adverbs and adjectives. At the moment we just don't have the lexical information to decide between intersective, subsective, etc. It's NON-specific. Underspecification implies some kind of forethought that you can't make up your mind. Non-specific is just "God, I don't know!"
oe: Here would the resultative have the same representation?
Dick: You might get a slightly different structure because of the parse, but certainly not anything like a resultative meaning.
Alex: But maybe later in the pipeline...
Dick: I'll see if I can connect.
oe: What about the resultative in the AMR?
Tim: I think it would be just the second one, not specially treated. We don't do resultative things that well anyway. If AMR was able to do it, it would be something like the thing that was the quantity of the votes.
oe: If vote had an additional argument, then it could be :Arg1-of.
Tim: Surprised-by thing, quant-of ... I don't think annotators would be allowed to do that.
Alex: I thought the annotators were instructed not to put in roles that aren't linguistically realized.
Tim: That's what I meant by not allowed.
Woodley: Why is it untenable to have surprise in at least a few of these cases take the event as an argument instead of a proposition?
Dan: If it's the event, you can't put in negation.
Ann: You can't really negate the nominal one: Their early non-voting surprised me.
Dan: But there you have to have an instance, because you're doing nouny things and the event won't be visible by the time you get to surprise.
Ann: The distinction between the resultative and the non-resultative reading with the nominal gerund is a distinction between the thing that nominal denotes, but it's a nominal type in thing in both cases.
Woodley: The fact that you can quantify it and modify it doesn't exclude it being an event. You can quantify events in some frameworks.
Ann: It depends on whether you treat the event that we have, the e, really denotes the event or is simply a convenience. There's nothing that says that the xs can't denote events.
Dick: Confirming no difference between their voting early and their early voting
Emily: So what's interesting about nominalization?
Dan: Staying away from gerunds now...
oe: Can avoid  (gerund) and the nominalized question . Looks now like we have some clausal complements and three examples with deverbal nouns.
oe: How did we come to this grouping?
Emily: The first three reflect where the ERG uses nominalization_rel, and the others came from the survey.
Dan: I think the 502 and 503 we can skip. The first one is a little interesting because there's a clausal subject that can be marked with either a to or a for ... to, though I don't know there's anything interesting to add about clausal subjects.
Emily: What I'm interested in is the asymmetry between subject and object clausal arguments, nominalized in subject case only.
Alex/Dan: Don't bind the barker here.
oe: Right---isn't that obvious?
Dan: Someone else might do it, like AMR.
Tim: Yes, AMR did it.
Alex: But that doesn't mean he put it there because he thinks it's syntax.
Tim: Right, I read an example that suggested that Browne was dog-like... There's an argument that I shouldn't put it in because they're sentences out of context, but I just made the assumption that Browne did it. I don't think I'd ever say this sentence unless Browne is the one barking. Which is probably more frequency than anything else.
Dan: You reckon it's by far the most likely interpretation. And then in a sentence like To bark is difficult you're not bothered by leaving the ARG0 of bark unbound because there's no binder there.
Ann: Do these things show up in corpora?
oe: We could probably query DeepBank for that, with the right fingerprints.
Emily: So what's the generalization behind the asymmetry again? Is it subjects that get nominalized, or ARG1 or?
Dan: That John sings bothered Browne/It bothered Browne that John sang. Doing that normalization causes some other desired normalizations to not work as well.
Alex: Can you talk about the asymmetry, because I wasn't at these meetings.
Dan: Let's first get the observe asymmetry: In clausal complements, I put the clause in as a scopal argument directly of the embedded predicate. I uniformly don't do that for subjects. There is a generalization that subject to pragmatic constraints, I can put a clausal subject in to any verb. The verb needs to be able to say that it's going to grab the relevant semantic element---would be INDEX for NP subject and LTOP for clausal subject. Put in the nominalization to simplify this. We don't have a notion of a GRAB-ME feature in the semantic composition that says which piece of the semantics an embedded context should pick up. I would need to split the subject-head rules into the two parts just as I have done for modification. For complementation it's all lexically driven. But my current view is that that's not true about subjects---it's not an idiosyncratic property of a lexical entry that it says I can, surprisingly, take a clausal subject. In order to accommodate this property of subjects, I've chosen to keep the grammar simpler and coerce these into nominalization. I have in the past played with a different approach where the subject-head rule is split into two types depending on the semantic shape of the subject.
Alex: So _bother_v doesn't take a handle for ARG1 ever?
oe: That the dog barks bothers Browne.?
Alex: So what happens in the passive? John is bothered that the dog barks. A scopal argument? Is it still _bother_v_1_rel?
Dan: Yeah, probably the same.
Dan: I do that uniformly for the verbs that can either take an NP complement or a clausal complement. I call it `p' underspecification between x' and h'.
Ann: We came to the conclusion that there's no point in differentiating in the grammar because it's trivial to do later.
oe: What are the two examples we should contrast?
Dan: It bothers me that Browne barks./That Browne barks bothers me.
Emily: No, because of your intraposition analysis, they get the same.
Dan: So this isn't the example.
Emily: Even if there isn't a minimal pair. The asymmetry is between sentential subject coercion via nominalization and clausal complements.
Dan: Currently, I'm treating only wh clauses as subjects; the only that clauses as subjects come from intraposition.
Woodley: You're not even allowing That zebras appeared causes me to wonder.
Emily: So when you have a wh complement of wonder, that one is not nominalized.
Dan: Correct. Different from the lack of idiosyncrasy of subjects.
Emily: But now you're saying that that CP subjects are only there by intraposition.
Dan: And yes that's an idiosyncratic lexical property.
Emily: So the any clause can be a subject generalization is just wh clauses?
Dan: Yes. And any wh clauses can be an object of a P.
Woodley: So For Abrams to bark bothers Browne gets nominalized and That Abrams barks bothers Browne doesn't.
[Testing with ERG]
Dan: Nope --- because of intraposition there.
Emily: But we're seeing nominalization in To bark bothers Browne
oe: It bothers Browne to bark?
Dan: This is something to be fixed in the grammar. Our ERS for this example is not write.
Ann: If you just took out the nominalization_rel as an end user, could you get a sensible semantics?
Alex: And replace the xs with hs in the appropriate places.
Ann: Would I be happy?
Alex: I think I would be happy.
Dan/oe: Let's look at some other representations.
Johan: I would like to withdraw mine --- sent new ones. I think the clausal complement should be a proposition.
oe: Who's x3?
Johan: We don't know.
oe: It's an unexpressed agent.... that would be parallel to what the ERG would produce when taking out the nominalization.
oe: So you'd have Johan on your side then. At least today's Johan.
Dan: And this one would generalize to not to bark?
Johan: Yeah, negation would in here, another box...
oe: Putting in a proposition gives him an ability to put in a negation.
oe: AMR doesn't have that challenge of making a home for negation because you'll just make it a feature.
oe: So in a sense this would also be parallel ... well, it's hard to say. The difference in denotations/possible interpretations that corresponds to the nominaliztion is not represented in AMR.
Tim: It's non-specified, I guess?
Dick: Mine had a hopeless parse.
Silvie: With us, barking is just the actor of bother, and we don't care whether it's a noun of a verb.
oe: So I think you'd be in good company then.
Emily: That's not how I interpreted Ann's question. I think she was asking if Dan puts in the nominalization sometimes, and ERS consumers do a nominalization-ectomy and then use the result?
Emily/Ann/Alex: But do you always want to take it out, and how would you know?
Dan: For nominal gerunds, you definitely want the nominalization because you can quantify: No voting early
Alex: Some analyses want the nominalizations as antecedents, but they're wrong.
Ann: If I have a PhD student who asks if I can get rid of it? Well sometimes, but where?
oe: I told my PhD student always get rid of it.
[SSI to find examples]
Dan: For verbal gerunds I'm prepared to believe you wouldn't need the nominalization.
Ann: The quantification examples are relevant.
oe: Used in 17% of all DeepBank examples.
Dan: Yes, prominent in NN compound constructions.
Ann: Can you briefly remind us that you do/don't have an ambiguity with the dancing teacher
Dan: Absolutely have. In the NN compound example, not the role-playing one.
Emily: When it's the role-playing one, you're actually binding a role?
Dan: No, because I don't know which role.
Dan: Let's move to the deverbal nouns, since this is getting too MRS-specific.
oe: Johan, do you ever do something parallel to this nominalization? Like when they're quantified. All early voting from Hull surprised the reporters.
Johan: That wouldn't be propositional. The discourse referent would be the argument.
Dan: No voting event in there?
Johan: The discourse reference can be an event.
Dan: What about no early voting would be allowed
oe: So you're using that polymorphism to its fullest.
Ann: But the lexeme would be voting, not vote.
oe: Not sure that's deliberate. It's what the stemmer is currently doing.
Alex: Where you've got polymorphic bother, he's got polymorphic voting. Is that right Johan?
Johan: That depends on the POS tag, which helps me distinguish between the two.
oe: Let's follow Dan's suggestion and move on from the one type of nominalization to a different type. Deverbal nouns, more interesting I'm sure to many.
Emily: Who's making the linking of the possessive specific and who's making it non-specific?
oe: I guess AMR makes it specific.
Alex: Re DRS same as this morning, he's got the possessive, and the possessive its up to pragmatics to work out what of really means here.
oe: So Johan leaves it non-specific.
John: You don't specify that government is the agent.
Johan: Well, I could.
Alex: Surely you'd need selectional restrictions or something not very much to do with syntax to do that.
Dan: The government's abolishment/The government's abolishment of the law.
oe: You need a human in the loop, with world knowledge and context.
Johan: Just thematic role labeling. The structure would be the same, just the name of the role differs, and that depends on your inventory.
Dan: It may not fill a role: Today's repeal
Alex: A FrameNet role that kind of mixes arguments and adjuncts. That's another reason to have roles extend to things like spatiotemporal locations, this is outside the grammar. Like with real possessives like John's book.
oe: Sylvie makes this actor/patient.
oe: So you're with AMR on this one.
Silvie: I know it's hard to do it automatically, but as an annotator, I see it so I put it in.
Dan: For Wednesday's repeal, you wouldn't try to put it in.
Silvie: It would still depend on repeal, but it would get time-when or something like that.
oe: [Calls up xfr] That's rich.
Dick: Reduce the font size...
Dick: There's a bunch of roles because there's a bunch of deverbals: government, voters, repeal. There may have been a voting by the voters, because you get that from the deverbal.
oe: Exactly where does that come from out of curiosity? The knowledge about these derived forms.
Dick: You can produce a list of -er nominalizations.
oe: So in the grammar's lexicon?
Dick: In the semantic lexicon. The grammar doesn't care whether a voter has a derivational link to vote. But this is useful semantic information.
Alex: What does your semantic lexicon feed off? The original word forms or the predicate symbols output by the grammar?
Dick: The latter, but for that's going to be citation form/lemma. There's a variety of things that go into this lexicon. We pulled data from various places, but not all of it gets used. Verb thematic role mappings, which we don't use.
Alex: But you did for plan yesterday.
Dick: We don't actually use those. Part of the lexicon goes through looking for various alternation classes.
Emily: Like the Levin classes?
Dick: Certainly not all of the Levin classes. Going through the data, there were only a few that I was confident enough were real alternations. Rather than using VerbNet role names for that, we have a notion of logical subject, object. Currently for the CLEAR parser + XFR, we have access to Kathy Dahlgram Cognition with naive semantic axioms, like pilots fly planes. John is a pilot. Does John fly planes? The deverbals (paper with Olya Gurevich)---she did a lot of work do pull out all nouns that seem to be related to verbs and then classified them according to patterns in behavior in compound nouns, possessives, etc. Four or five way classification.
Alex: But that kind of info in compounds is defeasible at the discourse context level.
Alex: Is it possible to recognize that defeasibility at the output of your semantic processing.
Dick: Yeah, and if you're asking for that trace, don't complain at how detailed these are. This thing tends to be a bit of a pack-rat in terms of trying to keep hold of that info for derivation.
Dick: The semantic lexicon is kind of a hodge podge of things all of which are really useful.
Alex: ...for doing pragmatics along the line.
Dick: I have a hard time saying where the dividing line between the semantics and pragmatics is going to be, partially because of the implementation. History: transfer semantics kind of came about by accident. At PARC there was a project to connect parsing up to ontologies and reasoning. At that point it was glue semantics, and there was some sort of baroque truth-maintenance system trying to deal with that. And I had been told to reimplement the transfer system to handle ambiguity better and that turned out to be useful for the reasoning task. The glue semantics froze because it was good enough to feed in more than enough problems. But people complained "I don't want to know about the linear logic" and I was too busy not doing the semantics. You're by now used to using the transfer system, and I can create the semantics that way. All the way to the back-end system it's transfer rules, so it's hard to tell the boundary between semantics and domain-reasoning.
Alex: Does that matter though? What matters for me is that info that you are suggesting about context that is likely but doesn't hold everywhere, I can trace it back and work that out. I need to be able to distinguish the massively likely because of domain knowledge and lex semantics from dead cert.
Dick: Like tense/aspect done by Cleo "in the KR"... it was a political not technical/theoretical distinction.
Alex: Aktionsart (Munson/Steedman paper)---you can change run into an event in an appropriate discourse context: Max ran in four minutes this morning if the context supports a regular distance interpretation. That's not the linguistic signal.
Dick: That's why I'm so down on compositionality now. You're driven by the structure of the parse.
Alex: in four minutes is looking for an event, and that is compositional, but marking run as an event is not.
Dick: I still think of that as part of the semantics. There are a whole load of things that are not strictly driven by the structure of the parse but are still semantics.
Alex: With your veridicality stuff, I wouldn't even be working this out at the sentence level. Knowing what the triggers are, I wouldn't compose that stuff in the grammar. I'd have that in the lexical semantic module.
Dick: There is a set of rules/algorithm for how you'd do the veridicalities, but to set up a purely compositional semantics for it would be a nightmare. I'm not sure Cleo would be so happy to drop compositionality, but when she was working on e.g. present perfect interpretation there were constraints, but it wasn't clear how to do it compositionally. No one further down is going to want to touch the representation you'd create, without a good deal of further simplification. Generalization to the worst case makes the whole thing unusable for your customers. Easy to state as non-compositional rewrite rules.
oe: Back to this representation...
Dick: Anything with govern_v/repeal_v/vote_v could cut out those lines. An a really expert emacs user could do that in a few keystrokes. If you cut through all that, like Silvie, we were filling in all the roles. The one thing that puzzles me a little bit is that this representation says that the repeal happened, but I'm not sure that that's the case.
Dick: To actually get any kind of non-veridicals, in this set up, you've got to introduce a new context.
Alex: Where's the veridicality?
Dick: Everything is in context t, except the voters vote. But I'm not sure what would trigger the embedded context. Johan's representation was the same.
Alex: Maybe nominaliztion is the embedding.
Emily: But this is a deverbal noun are we going to call that nominalization?
Dan: Did you introduce an embedded context for the gerunds?
Dick: The -er style deverbals that is really set up to introduce this. Some of the other ones, although there's an implicit governing event, are set up differently because it's so rarely needed. But even putting those in would not help with this case.
Dan: But again back to the gerund form: the government's repealing of the law would that put in the context because of the gerund form?
Dick: No, the gerund wouldn't give you that context.
Alex: Maybe you could finesse it because you are in an intensional framework, and with support and because these refer to concepts not denotations there.
Dick: The standard intensional verbs that actually take a nominal argument (seek, prevent, support) need to put in context lexically, since it's not automatic for noun complements.
oe: Are you resolving the roles?
Dick: I'm being non-specific because it's a genitive.
Dan: But you're also putting in the object relation.
oe: How did you know that that was object and not subject? The voter support the law's repeal As in, how does your system?
Dan: The government's repeal --- what automation are you going to use to decide that one of those fills the subject role and one fills the object role?
Dick: It's going to be in the semantic lexicon.
Alex: Surely it's a Kathy Dahlgram thing, that governments repeal laws.
Dick: This doesn't have the Cognition stuff in. Give me a while to dig out the DB and see what it's got for repeal.
Tim: We also do the govern and vote form voters and governments.
Dan: And you worked out that the government is doing the repealing and the law is getting repealed in both cases. And if you had the government's appeal?
Emily: Then he would have looked at the context and decided what's appropriate.
Tim: government-organization: That goes to an organization that actually happens to govern.
oe: And a voter is a person who votes.
Dan: How come the repeal isn't represented as the result of the act of repealing. It's a deverbal noun too, but you didn't pull it apart the same way?
Tim: That's a design decision with AMR, a little problematic, didn't decide to split off the two readings.
oe: The voters support the act of repealing.
Emily: You can't tell here. The voters could support the act or the result, and it should be the same in the same situations.
Dick: So about repeal: The lexicon says if you've got a genitive, if you've got a possessive, and compound-noun constructions what's it going to map onto. For repeal it says both the genitive and the possessive map onto the object, but when you've got both one goes to subject and one goes to the object.
Dan: So the presence of that of pushes you away from what you would otherwise do.
oe/Dan: So then you got lucky with how we constructed the examples: The government's appeal you'd treat as object only.
Dick: Hard to compile this data if you're also going to try to catch the corner cases.
oe: For the ERS: the repeal is a noun in the lexicon, as are voters. And the government is non-specifically linked with poss_rel, but the law is definitely the object. And we have at times wondered about whether we should make the underlying verbal argument structure overt.
Dan: That's part of what's interesting to me. It's a rode we have not gone down, but AMR does really aggressively and it looks like Dick does that too. Do you do that systematically for other things too Johan?
Johan: I assume that there's some role labeling that does that.
Dan: Do you systematically expose the verb's relation when it's not a gerund, in either productive of fossilized derivation?
Johan: I'd like to have the agent nouns, and e.g. victim as patient of some crime, but I don't have that now.
Ann: It doesn't seem to make any sense to me to do it just for the -er ones. What about tutor where it's clearly teach, and clearly not tut.
Dan: Not really teach either.
Ann: It's to tutor. There's something clearly agentive going on anyway. On the other hand, with some of the -er nominalizations, it's not the general verb. I don't really see a reason to put the established ones ... I think it's a cline.
Johan: I want to have this because it's useful for logical inference. The voter is the victim --- voter(x) and victim(x), you get an ontology clash.
Emily: That depends on your analysis of the identity copula too.
Johan: Yeah. What I'd like to add is something that says person(x), so there isn't a clash.
Dan: Your enemy is my friend --- does that clash?
Emily: He wants them to both be expicitly persons.
Dan: Aren't voters and victims both persons?
oe: To Johan making that explicit is the same thing as decomposing the agent role.
Johan: In my automatically generated bg knowledge, I have info like a voter is not a teacher.
Dick: Most ontologies do not do a very good job of specifying disjointness.
Johan: For me teacher and voter is something like a role. Not just agent nouns, because victim is a patient now. I'd like to have an ontology that says these are roles a noun can play, not the person itself.
Dan: Will that scale to friend or enemy (no verb there)
Johan: Those are relational nouns---so there's a role there, too.
Dan: I think you're hoping to get to persons because that entity is a role-player in an event and can be free to be a person. I don't think that's the generalization.
Dick: You're automatically trying to generate disjointness assertions, based on shape of the hierarchy. voters are people, teachers are people: algorithm says those are disjoint. Role relations would be a way to say those aren't disjoint. But there are probably other things where that won't help. Just synsets under the same hyponym doesn't mean they're disjoint. Not just lexicographers but also most ontologists are really bad at putting in disjointness information.
Johan: I need the disjointness. It's very important for me. But I'm still looking for an example.
oe: For me, being tall, the best chairs are tables.
Dan: Any pet is a burden.
Ann: burden is relational...
Ann: You can get the copula between singular and plural A rugby union team is 15 players.
Dan: These guys are an utter failure.
Johan: We're going to discuss the copula this afternoon?
Alex: Let's hold that thought for
oe: Why did we include the student evaluation examples?
Emily/oe: It was an AMR trap.
oe: For those folks who believe they have access to competent speakers of English.
Tim: Went back and forth, but decided it was the students doing the evaluations. [Actually provided two annotations, one for each.]
Emily: Not leaving it non-specific.
Tim: Not allowed.
oe: person who studies isn't necessarily all there is to the meaning of student.
Tim: Same problem with voter: a likely voter might not show up at the polls, but they're still a potential voter...
oe: And ARG2-of ... that's the result?
Tim: That's the result not the activity. [corrected below]
Dan: The evaluation never ended.
Tim: We have a role within evaluate that is the conclusion. But that's different from the more general sense of the result of evaluation.
oe: A three place argument, where negative is the third argument (arg2) of evaluate.
Emily/Tim: The evaluation never ended. the argument of end is (e / evaluate).
Silvie: evaluation is not shown as an event, so student is RSTR. If it were treated as an event, student would have been ACT. Without something that recognizes the event nouns it's not treated that way.
oe: You think this should be ACT?
Silvie: Oh, I remember. I did RSTR as a way to leave it non-specific. There's no clean way for me to say it can be either or.
Dick: You do get student put in as the object of the evaluate verb. I think the subject reading isn't very prominent.
Emily: Have you never taught a course where you were subject to that?
Dick: Semantic lexicon says that in compound nouns, the modifier generally goes to the object.
Alex: You're being too specific there.
oe: Dick with his deep lexical pockets is the least non-specific in this example, possibly to a point of being too specific.
Alex: What is amod doing there?
Dick: Nothing very fancy is being done with adjectives. Given that you have both amod on the verb 'evaluate' and the noun 'evaluation' suggests that it's the process that was negative.
Alex: You're not sneaking in two negations then.
Dan: No, that's the real word negative.
Ann: Can we just talk about something sensible or go eat something?
oe: Let's go eat then.
=== Ellipsis ===
Emily: So it's coordination or ellipsis right now.
oe: Is it?
Woodley: Let's start with the AMR of the first example.
Tim: These are dreadful for AMRs.
Emily: Because we gave you no context?
Tim: Yeah, because we have zero cotnext. AMR is supposed to be somewaht grounded in terms of what you can introcue, and there's no concepts here. The most bland things in our ontology is individual or event. AMR-unknown is just an interrogative component. When Who is? is asking who is doing that action, it's event-01 with an ARG0 for the wh. For equative clauses, then it's a nominal relationship, expressed by thing.
Woodley: You could have also represented the first version as :ARG0-of if you wanted to emphasize the person.
Emily: The AMR guidelines say that that represents focus.
Tim: ... but we're really vague about what focus means. In practice, it's usually the main verb or discourse connective.
Johan: Is the personhood of the who encoded?
Emily: Should it be?
Tim: Abstractly, yes, but in practice no.
Tim: All you're getting is something to hang your coreference links on.
Emily: But you don't do it, because it's cross sentence.
Tim: For now...
oe: Yes, Johan, ellipsis.
Johan: They're not there...
Dan/Woodley: Pretty good analysis.
oe: You're a man of principle.
Johan: The recovered material should be, but if you don't give me context, I can't give you the meaning representation.
Alex: That's because you don't have underspecification.
Johan: I look forward to your underspecification of this.
Alex: David Schlangen's thesis.
Alex: It's perfectly possible to have multiple derivations of Abrams is, with underspecified predicates in all of them but hte copula doing different things.
Johan: But what does 'Who is?' mean without context.
Alex: But it appears in dialogue. If your grammar is producing a logical form, and it needs that recovery before it can choose waht the logical form is, then I don't understand your architecture.
Johan: Here it's syntactic ellipsis.
Alex: No: Where are you headed?/To buy an apple.
Emily: That's not ellipsis of this kind. Does it matter?
Alex: No. There's this huge controversy in the literature over whether you can recover ellipsis by copying syntactic structure and I think the current consensus is tha tyou can't.
Johan: I can for VP ellipsis.
Alex: It's not entirely syntactic because of sloppy and strict.
Johan: Sometimes you copy, sometimes you don't.
Alex: When you don't copy bits of syntax, how are you going to do it? I agree it's hard to do, but you seem to be arguing you can resolve the whole meaning within the grammar if context gives you enough, but I'm saying that's the wrong place.
Johan: If this is your preliminary representaiton that's fine.
Alex: That's what underspecifying is doing.
Johan: Maybe. Some people also think you can give an interpretation to it.
Alex: As in model-theoretic interpretations, where the models are what's going on in the domain? That's kind of mad.
Johan: I could give you the lexical entry for Abrams, and four different derviations with the different forms of is and even combine them with CCG, but that's not the final interpretation.
Dan: The first three seem fairly messy, I would have thought you could do something with the fourth one. You know where the missing event/proposition fits in, and that Abrams is related to it, even before you see the discourse. I was hoping to see something about the division of labor, what you can do in the absence of context. For Johan it sounds like that's not a very interesting exercise because that stage of representaiton in the pipeline isn't very interesting.
Johan: It would be if you tell me how to resolve these things.
oe: So you suggested to Johan that he could have constructed something similar to what we have in the ERS (on the screen now)?
Dan: The role of Abrams in the embedded proposition has to be the external argument...
Emily: But that's syntactic.
Dan: Yes, because the ellided VP could have been passive.
Dan: On my view, this is providing some relevant information for some NLP system, but it's not doing a large part of the work. I know where that eventuality should go, what it's shape is, something about its scope constraints. In this case I'm with Johan, you do care about the shape of the syntax of the antecedent.
Emily: That's kind of scary.
Dan: Everything about ellipsis, is.
Woodley: Who was eaten?/Abrams tried to
oe: Who tried to be eaten?/Abrams tried to
Emily: That's the controversy Alex was alluding to. There seems to be some syntactic constraints, but there are cases where it can't be syntactic copying.
Alex: What do you do with me too?
Dan: There's a me that fills some role in some underspecified eventuality, which is modified by too. Two fragment elements both linked to an implicit eventuality.
Johan: Can I see Dan's who is?
Dan: I on this view don't tell you whether that was Kim is happy, Kim is singing
Johan: Or Kim is a woman
Emily: Definitely not that one.
Dan: Why not? Because be a woman could be what the ellpisis_ref_rel resolves to. ellipsis_ref_rel says there is some eventuality supplied by the discourse and Kim plays some role. Named ARG1 here, but that doesn't mean it has to be ARG1 in the resolved role.
Ann: Has anyone used these things after David?
Dan: Did you (Woodley) need them in the robot stuff?
Alex: In the robot and settlers thing I'm doing we're using different grammars.
oe: For there were we left two analyses. The one case where we didn't fully resolve.
Alex: THe one place where we are using these underspecified predicate symbols a lot is in gesture, where the form of the hand tells you something about meaning, but very little. Depending on how the gesture relates to the linguistic form that its cotemporal with, you do get some info about what part of the linguistic form comes into the meaning of the gesture...the equivalent of that x5.
Woodley: For there were can you explain why ARG1 is an event?
Dan: What we can say most positively about the interaction of ellipsis and presentational there is that the grammar has made a decision.
Emily: Some thought went into this, because otherwise you wouldn't have ellipsis_expl_rel.
Dan: What I can't explain is why the ARG1 is an event... it should be an ellided individual.
oe: Take that one off the table and move to the next one.
Dick: The XLE parser doesn't do a really good job parsing the elliptical sentences. It's introducing some null verb and claiming that the who is the subject of that. My guess is that this is because the parser has decided this is the auxiliary is rather than the copula is.
Dan: If it had been the identity copula, it would have been the copula in there instead of the null verb?
Dick: Yes, I think so.
Alex: Question is veridical?
Dick: You introduce something that is an interrogative context right at the top.
Woodley: Do you ever have a situation where the top-level context is not veridical?
Dick: Only in a discourse context...
Dan: In the case of did you I'm quite surprised that you have a do.
Dick: A weird parse as well---not a question, and not the auxiliary do. And it might be the only.
Emily: I'm very curious to see what you do with try to because of the veridicality.
Emily: What are you doing?
oe: I'm taking out the crap.
Emily: What did you take out?
oe: The crap.
Dick: No, you missed some.
Dan: So the subject is linked to the subject role of the null verb.
oe: This is logical subject.
Dan: That doesn't mean it's the logical subject of the resolved predicate.
Emily: The null verb is averidical... that comes from try.
Tim: AMR just doesn't have the null verb there. Just a trying situation. All null instantiations, with no antecedent in the sentence, are left out.
Dan: Here try some fish. *No thanks, I already tried. So transitive try is obligatorily transitive.
Tim: Leaving a race: Well, I tried.
Dan: Tried to do something.
Alex: With the try that takes a noun v. the one that takes a VP, is that one of your polymorphic guys.
Dan: I probably make it polymorphic, but I suspect it's different senses.
Alex: When it appears with ellipsis, you should resolve the type of the ARG2 because it can't be x.
Dan/Emily: The ERG does, because ellipsis_rel is the ARG2. ... well for Abrams tried to but the grammar is currently wrong in treating Abrams tried as argument drop and not ellipsis. No ellipsis_rel, ARG2 is p.
Dan: Curiously, I attempted isn't good, even though I tried is fine, but has to have the VP complement reading. There should be a class of verbs like try that take a VP complement that do license VP ellipsis.
Woodley: And NP drop.
Emily: That could depend on the verb. And if there were a verb that could take either an NP or a VP and drop either, you'd need separate lexical entries.
Alex: Abrams tried parsed with ARG2 p ... something should say that p can't be a noun thing.
oe: Dan intends the MRS for Abrams tried to be the same as Abrams tried to.
Alex: I don't want the grammar overgenerating the possible readings.
Emily: If the grammar should rule something out, you want it ruled out.
Emily: But if try actually allowed drop in either the NP or VP case, would you be happy with [ARG2 p]?
Alex: Yes, because then I can decide.
Alex: When the grammar's overgenerating, the discourse representations get overgenerated as well, and spurious ones can go up the ranking.
Silvie: We have #VerbPron for a coref marker that necessarily goes back to some linguistic antecedent. We can have exophora, but this would not be the case here, since we know the verb from the context. When we don't have any context, I would just make the antecedent missing. I picked bark as some random verb, because usually you annotate text with context. Like all others, we are bad at annotating things without context.
Alex: Where's the question bit?
Silvie: Within the inner structure of the node.
Woodley: It seems to me that although the point about context is an important one, if you want to have a theory under which the meaning of discourse is compositional, in order to compose the meaning of the discourse where there's parts, you need the meanings of the parts first.
Alex: That's what I was trying to say for Johan.
Woodley: Even if it's not a fully-formed meaning...
Alex: But it's not a free-for-all either. It can't mean Fido. It's very useful to get the grammar to tell you what needs to be plugged to make the thing interpretable at all. There's a gap that needs to be filled, and you can't interpret the discourse at all until you fill it. This is about construction of the logical form that will give you an interpretation.
Woodley: Whereas in a theory where you're not trying to do things compositionally.
Alex: Dynamic syntax does it this way. Syntax transcending sentence boundaries, give a syntax for the entire dialogue that way.
Johan: I got an annotated corpus of VP ellipsis---the whole WSJ---if you want to test your ellipsis component. www.let.rug.nl/bos/vpe/corpus Looking at tv antecedent case.
Do you think the British know something that we don't?
Johan: Do you link the something to ellipsis_rel as well?
Dan: I don't even parse that. Happens in other modals: He does something that I can't. The British will say things that I won't. There's a gap that's got to be filled, but there's no place for a gap. That would be a fancier construction with a missing VP with a hole in it
Johan: I tried to organize a SemEval task around this, but no one seemed interested. Must be because they can't do it.
Alex: I'd imagine it's massively ubiquitous in dialogue data.
Dan: Even if in boring text it's pretty frequent.
Johan: Really common is comparative ellipsis. "and X Ved too" is very rare. Also sloppy/strict. Only 10 cases of potential sloppy/strict ambiguities, 9 were sloppy and one was strict. Then we tried a larger corpus, and still 95% sloppy.
Woodley: You more or less declined an analysis of out of context ellipsis, but you've done a lot of analysis of in-context ellipsis. Can we see one?
Johan: It wouldn't look elliptical.
Emily: There seems to be something to say about which levels of representation are interesting. Is the underspecified grammar output interesting?
Emily: Woodley said before that discourse meaning is compositional, and you didn't blink.
Alex: I did, but quietly. The combination of the various bits is not strictly compositional because it involves defeasible inference. It's like global decoding. If you add new sentences to the discourse prior values/ways of resolving stuff for prior sentences can change. But I do want as input little bit of meaning representations. Not like Dynamic Syntax/DL-TAG.
Johan: I think it works for some cases but not all of them.
Alex: If it doesn't work for all of them, you may as well work off the meanings, and make sure that the meanings carry a trace of what you could have done syntactically. That the meanings are sufficiently rich so that you can essentially simulate that syntactic copying but with the semantics, because sometimes it is semantic.
Johan: Andrew Kay says it depends on the discourse relation whether it's syntactic or semantic recovery.
Alex: And you won't know if it's parallel or contrast without knowing the semantics. But to identify that you're dealing with parallel or with contrast, is about reasoning about the meanings of individual things and how they relate to each other.
Johan: If we lose all the syntactic information, you'd lose a lot of clues to find the right antecedent.
Alex: I'm still hopeful that if syntax rules in/out certain things will be enough. Will have the info in it that you can attach that constraint to in discourse processing. I'm yet to be convinced that that's wrong.
Woodley: On that compositional question Johan you said you can show us resolved DRSs, but not the intermediate pieces. How do you build the resolved form? Manually? with a mechanism?
Johan: Yeah, I can show them. gmb.let.rug.nl/webdemo/demo
Every man loves a woman.; run the pipeline, check syntax, semantics, then expand the derivation to see CCG and DRT bit.
Alex: So you're doing higher-order unification on the lambda expressions...
Woodley: The question is how you're handling the resolution of VP ellipsis.
Johan: I have something but it's experimental code, not integrated yet. Basically what Nicholas Asher did. Abstraction is applied to the missing stuff.
Johan: If you try Every man does.
oe: Then you have to do something; we have to make you do something.
Johan: This is my representation, but I'm not happy with it.
Alex: You'd need an interface to lexical semantics that says that do is not intransitive, might is not intransitive. POS tagging puts intrans verb.
Dan: But POS taggers do discriminate between intransitives and auxiliaries.
oe/Woodley: But the CCG tag is intransitive verb.
Dan: I thought that taggers were smarter than that.
Johan: But C&C has to get a derivation, and this is the only way.
Johan: I don't know how to do ellipsis really. I can hack it, but not really linguistically motivated. Can spot modals treated as intransitives and then apply something to that. Haven't seen any complete computational implementation of VP ellipsis. Plenty of theoretical work, but look at the variety of the examples in the corpus.
Dan: If you tell me it's impossible, I won't waste any time on it for a while.
oe: We've gained new respect for the phenomenon, seen an excellent new resource.
oe: Who's going to chair?
Emily: I nominate Alex K. because he brought it up.
Alex K: What?!! What's involved in chairing?
Emily: You keep the discussion going.
Alex: You shut people up.
Alex K: Let's start with the one on the screen.
Johan: There are two referents, and I think they should be identical, but if you just do x = y, you can't attach any modifiers or tense. So I need a trick: I put them in a possible world/a situation. Don't really like it, but don't see a better solution.
Alex K: What problem?
Johan: Distinguishing between Abrams is/was/used to be/will be Browne.
oe: One of these two boxes is there to have a place for the tense information.
Dan: might be Browne
Johan: Then there's be a box for the modality.
Emily: Just one?
Johan: No then you'd have two, because it's compositional.
Emily: But to oe's question... why do you need a second box for the tense? Why can't it just hang on the outside one?
Johan: You need a discourse referent for the tense to hang on. Here p1.
Johan: On the equality representation: I tried to find a compromise between linguistically adequate and logically sensible.
Dan: The small price you pay is that you get two boxes with the modals?
Johan: No. The modality operator doesn't introduce a discourse referent. In Boxer there is also an option 'copula true or false' if you don't like this and would prefer a being event.
Johan: Copula adjective, copula PP aren't clearly the same. But it's the same problem. Or maybe you don't have that problem. Abrams is old still need something to put the tense on.
Dan: And you're not supplying any kind of situation for the adjectives.
Johan: I do now, since a year. But I don't for the PP. Probably have to make that a three place relation. But I don't like it so much.
Dan: Abrams is present v. Abrams is here will be quite different.
Johan: No. In CCG that will be the same category...
oe: Abrams is in the room?
Dick: It gets to be really weird, prepositions having events if they are modifying events.
Alex: Not very Davidsonian.
Alex: You can get rid of the event if it's giving you trouble.
oe: In constructing these examples, we didn't include what we call the predicative copula.
Emily: Because it's not in the semantics in our universe.
Johan: What do you do with the preposition.
Ann: I think you should say eventuality, not event.
Dan/oe: Right, that's what you mean.
Johan: He was walking about in the room --- you have an extra event there.
Dan: Yeah, but it can be inert.
Alex: It's annoying to have it there, which is why I said post-processing you can get rid of it.
Johan: But in your ontology you don't want in sometimes two place, sometimes three place.
Dan: What do you do with the adjectives?
Johan: Always two-place (always with the extra argument slot).
Dan: But you're not ready to do the very same thing for the PPs. The dog angry at the cat v. The dog in the park?
Johan: But that's a selected PP by angry
Dan: Okay, the dog happy yesterday: The position of the intersective modifier isn't interesting. Some adjectives are post-nominal, if they end up phrasal.
Dick: Is the additional eventuality only to hang tense off of?
Dan: It supplies the convenient piece of combinatorics for modifiers: right under the table, just in the park
Alex: The dog in the park yesterday was out with my daughter last week. yesterday has to modify the in eventuality.
Alex K: It seems like the predicative copula is also interesting topic of discussion.
Cleo: Isn't Browne is a manager also predicative?
oe: We consider that a borderline case. We were just discussing why we did not include adjective or PP examples which we consider a predicative copula.
Alex: It is an identity---because you've got a quantifier there.
Dan/Cleo: It may not be a quantified expression. It may be a property.
oe: We included it because we expected it to be a borderline case.
Alex K: Should we ask people to give their opinions about the predicative one? Johan your position is that in Abrams is old ...
Johan: Go to webdemo: there's a being old state (s1) where the tense is. but haven't added it for PPs, so there's an extra box (p1).
Alex: You should just bite the bullet and put in an event in that preposition, and you're troubles are over.
Johan: But just open any textbook on formal semantics...
Alex: But that's because they don't think about these things.
Dick: buttered the toast in the room There's an event, and then a prepositional event that connects the two. If we're going to semantics rather than syntax, what the hell does that second event mean?
Alex: The event of buttering the toast is in the state of being the room.
Dick: Can it change?
Alex: It's got spatio-temporal extent.
Dan: frequently in the room --- what is frequently going to quantify if it doesn't have something to count. The dog frequently in the room
Dick: This is saying that all modification has to be done in a Davidsonian way. Predicate modifiers are completely out? Why is that? There are other ways of modifying modifiers.
Alex: If your logical form is higher typed like that, in terms of model theory, you're ...
Ann: We made a decision not to have predicate modifiers because of the availability of this ultra-Davidsonian representation. I think that thinking of these as events is wrong, but I think there's a gradation between the real event cases, and where these things are more ... not even situation-like. There's a series of papers that actually use that trick and the trouble is deciding where it's which is a slippery slope.
Dick: What this comes down to is every clause you're going to find yourself wanting to give that clause a label just in case something wants to modify it. Just seems really ugly.
Johan: Also what Jerry Hobbs is doing.
Dick: If you're not careful, you're begging all these self-referential paradoxes.
Alex: But not at this level, this is just a description.
Dick: That's like saying we don't care about the semantics of the semantics.
Alex: I'm not uncomfortable with that.
oe: Guilty as charged.
Alex K: You could just do a non-compositional post-processing step. Not saying that as a thing I would be happy with.
Ann: I think there is a bit of justification for this, but I got someway into looking into the literature about it, and then ran out of time. It's to do with these more situational type things---difference between pre-sentential and post-sentential adverbs. We are trying to do this in a way in which the semantics ends up similar for English & German, and in German it's much clearer that attributive adjectives can be temporally modified. Once you say that's going to happen, and adjectives and adverbs aren't that different...
Dick: Temporal modification of an adjective, does that pick out a time different from the actual temporal reference of the clause?
Ann: The once-green lawn is now brown.
Alex: Are there any languages where you get time info on prep?
Emily: Not that I know of, but in Chinese it's hard to draw a boundary between prepositions and verbs.
Alex K: Why don't you do this with the identity copula?
Johan: Because then I have a three-place equality relation.
Alex K: What's the difference between p1 and s1?
Johan: Just cosmetics.
Alex K: Then you still have a three-place relation here, you just don't see it.
Johan: Yeah. But I use this usually for the modal relation, not the copula.
Alex: And you wouldn't want frequently modifying possible worlds. The dog is frequently in the room.
Johan: It's a usual possible world or something like that.
Alex: Right. It's an abstract object, a discourse referent. Reified. Made salient. I'm not wildly happy with modifying that usually.
Dick: In something like old man, is it just man(x) and old(x)?
Johan: Not anymore. Otherwise you don't get inferences.
Dick: This is because the only thing that gets this kind of variable is going to be a discourse referent. The other question is if you've got this eventuality for the adjective, does that actually become something you could refer to pronouns?
Johan: In the formalism, yes. Linguistically, probably not.
Dick: Another way of doing this is ... scope referent for everything other than function words. But they refer to concepts, not individuals. For very old man, you'd actually have an amod from man to old, and then an amod of old to very. That's not the semantics introducing any kind of eventuality. You've just got the man. It's a little harder with PPs. I tried it at one point and backed out of it don't remember why. Right now, the prepositions are being treated more like function words: in(x,y). Used to have pmod and parg. That's carrying no ontological commitment to having these additional eventualities.
Dan: People in the room yesterday ... here yesterday is modifying the prepositional phrase. If I don't have the same shape for PPs as NOMs and verbs, then I've got to do something very different for the yesterday.
Dick: That's where the pmod would be useful...
Alex K: Let's look at the XFR representations.
Dick: Rather than introducing a being event, we're hanging the tense information off the context. Like what Johan is doing.
Alex: Except you're hanging it off the global thing.
Dick: After a little bit of simplification, ... I get a lower copula context that comes from XLE, but that upsets people so there's a rule that flattens that out and pushes the tense up but marks it on the context. I suspect that whenever you have something that carries tense, it introduces it own context and the tense information applies to the context, not the verb that's heading the clause.
Emily: So you see a strong difference between predicative and attributive cases.
Dan: So you're going to get very different representations for dog which is happy today v. dog happy today.
Dick: Think of contexts and world-time pairs. If you've got a new time, you should have a new context. So today should trigger a context.
Dan: The dog happy in the room/The dog which is happy in the room
Dick: Those are different, because you can also get The dog which was happy in the room. If you're worried about that and aren't worried about being compositional, you can wipe out the separate context for which is happy in the room to make it look like happy in the room. But that's not what you're going to get if you're actually driving it from the structure of the sentence. I think if you're being fairly compositional, you're going to get fairly different representations.
Dan: No I'm not. One just has more info about the tense on the eventuality.
Alex: But as Dick says, he doesn't like an 'in' relation between eventualities. It doesn't bother me so much, to be honest. I know no formal semanticists do it in their theory, but their models are kind of a joke, actually.
Tim: For the 'is in the room' thing, AMR does this reification game that sort of gets to this having a location relation and then having a variable for is in a room. We aggressively don't have ident relations because we say everything identical is the same variable. For Abrams is Browne we have one node with two names hanging on it.
Dan: What about Abrams isn't Browne.
many: Polarity -.
Dan: on what?
Tim: Could have a have-name relation.
Dick: What if it's not named things? The dog is the cat.
Tim: We have a loose use of the modifier relation, which I guess would probably be the closest thing which is really ugly.
Dan: And for pronouns? You are him.
Tim: Probably you mod him or you domain him? The model in which identity things become one node breaks down when you're trying to talk about things being identical things.
Alex K: Anything where you don't have direct reference to individuals, or the way in which you refer is going to be tough.
Tim: For things involving names, I use name as something about the identity. Abrams is Browne could be something like the identity of Abrams is the same as the identity of Browne, but we don't have a core identity relation.
Dan: Likewise for modals, like Abrams is probably Browne.
Tim: That's just a huge gap.
Dan: So the identity copula actually gets at one of the ground assumptions about how AMR is being done. Identity is so central to what you're doing...
Tim: If AMR expands beyond the sentence level will have to have identity relations, because lumping everything in the document into one node is just ridiculous.
Silvie: FGD displays the copula verb and can have tenses and all verbal characteristics there. The thing that is the individual is the ACT and the category name is the PAT. Had to assume that the word order is fixed, and that Abrams here is the individual being described.
Emily: And Abrams is old?
Silvie: That would be the same. Same label, same topology.
Woodley: Why no arrow between them?
Silvie: Convention. It would be redundant to the copular predicate.
Johan: Why do you want an arrow there?
oe: But in Abrams is old, I wouldn't expect to interpret that as identity.
Dan: You wouldn't be tempted to put in a blue arc there. So the be can't be doing it by itself.
Silvie: It wouldn't have occurred to me to make an arrow there, just couldn't quickly think of an argument.
oe: Do you ever get arrows between two non-generated nodes?
Silvie: Yes, with personal pronouns.
oe: But that's not #coref.
Silvie: Any occurrence of a pronoun becomes an occurrence of this lemma. Not necessarily a generated node.
Johan: (re ERS): So you basically stuck in the be. I can do that too.
Dan: There isn't much else I could have done. I don't have a way to put in the p1.
Johan: This is what you get when you say copula -. And so what about languages without copulas?
Emily: You have to have some rule of syntax that either promotes one NP to predicate or joins the two. That rule can supply the predicate.
oe: Brings up The theory is that Browne arrived.
Alex: Is this one polymorphic on the ARG2?
Dan: I put in a different predicate because it has a uniform shape with that propositional second argument. I was not comfortable just calling that the same identity. It seems strange to say that the theory is identical to the proposition.
Woodley: It should be the ARG2 of the theory also.
Dan: You think that's going to work all the way through? You're predicting then that there are no nouns that can be the subject here that can't take propositional complements?
Dan: Did anyone else do that?
Tim: AMR definitely being crazy did that. We verbalize things as much as possible, so theory and theorize is the semantic field.
Dan: The real disaster is that Browne arrived. Are you going to have disaster being a sentential complement taking noun?
Tim: I don't think we have arguments for disaster. It's just whatever arguments make sense.
Dan: So what would it look like?
Tim: We'd probably say that Browne arriving is modified by disaster.
Alex K: What would you say?
Dan: be_v_nv is a two-place relation between disaster and the proposition. I'm being the most skeptical about what you can associate with that structure, of attempts to do it in a squished fashion.
Woodley: I think that disaster does take a complement.
Emily: I think you're letting your theory shape your judgments.
oe: The disaster of John's arriving
Emily: That could be a modifier.
Alex K: John's arriving is a disaster ... why not say you're predicating disaster of it?
Dick: The disaster case convinced me that I don't want the more squashed representation.
Emily: *It would have been a disaster that Browne arrived. ... but that's probably just a tense problem.
Alex K: Is your point that you need a place to put tense?
Dan: I need a place to put tense, but that's not different from before with Abrams is Browne. The problem is that both disaster and the preposition are inert, not looking for any arguments. Need a relation to join them.
Woodley: I think that's right if you can come up with an example which is unequivocably non-relational, but we disagree on that for disaster.
Johan: DRS: additional box for x1 = p2 to put tense on p1, p2 is the theory.
oe: I interpret that as being the same as the ERS.
Alex K: Isn't x1 = p2 a type mismatch? What kind of object is p2?
Johan: A theory. I have only entities in my domain.
Emily: x, p, etc are only cosmetic.
Johan: Right. I just use some orthographic distinction, but it really depends on your background ontology what is what. p2 isn't necessarily a proposition. At any rate it's just an entity in my model. If it makes you happier, I could write p1 instead of x1.
Alex K: Maybe.
Johan: Would have the same analysis there, Dick?
Dick: I'd get almost exactly that.
Alex: So two contexts?
Alex: And would the embedded one be veridicality a?
Dick: I don't know. Just because it's a theory doesn't mean the theory is true. But it's nothing to do with the theory part of it, just that one half of the equality is a clause.
Woodley: The misconception is that Browne arrived.
Emily: The disaster is that Browne arrived.? Is the clause veridical there?
Dick: It feels like that to me now.
Alex: So we have n, a and p all dependent on the subject NP.
Woodley: You've walked into my trap.
Ann: I have never disagreed with you Woodley.
Alex: It would at least solve that.
Dan: The real bonus is that Browne is going to be at the party? Do you stop anywhere? Is there any noun that doesn't take a sentential complement?
Alex: The cherry on top is that Browne will be there.
Dick: It's not a sentential complement. It's that that noun applies to something. It's just what nouns and adjectives do---we don't say that adjectives have nouns as a complement. If you're prepared to allow free intermixing of regular objects and propositional objects, like you'll want for terms like theory and misconception, they're just applying and describing those objects.
Dan: It's just predicational.
Alex K: That Browne arrived is a disaster.
Dan: Will you take it to Abrams is Browne too? If you take the view that nouns can apply to whatever.
Alex K: There's an =Browne predicate that you're applying to Abrams.
Dick: No, more like Browne(Abrams).
Alex: Holy crap!
[Ann makes her exit.]
Tim: At least in AMR perspective, if you think of theory as a thing that is theorized, that thing can be identical to something, and theorized has scope over thing.
Dan: I'm not worried about that for theory, since it has a way of combining with a proposition.
Woodley: Dan you've entertained the idea that proper nouns are pretty close to common nouns, so Browne is a manager and Browne is Abrams can look similar.
Dick: Starting to remember some of our discussions about copulas. When you've got a copula predicative, you're taking the intersection of two concepts --- managers named Browne is non-empty. Is that what we were doing, Cleo?
Dick: And can you remember why we spent so long discussing that?
Cleo: I have to reconstruct it, but I think one of the issues had to do with what to do with the tense information coming in from the sentence.
Dick: Was there any problem with negation in that?
Cleo: I don't think the negation would have been a problem --- they don't intersect. I thought more of the problem would be the Abrams is Browne, because we couldn't get down to the individual level.
Dan: Let me help you with that bridge a little. You can certainly in talking about families say, John is obviously Browne/John is obviously a Browne --- setting up a set of individuals using that phrasing. It's still a jump to say I've got two singleton sets and make one predicate over the over.
Cleo: We didn't have easy resources to specify a single set.
Dan: Abrams is the only Browne: I've already told you they're singleton sets, but that's still impeccable.
Emily: Why is there a problem with intersecting singleton sets? That just means they're the same set.
Alex K: Are you asserting or presupposing that just now? Abrams is Browne it's presupposed.
Dan: Maybe that matters a lot.
Alex K: Names are weird.
Johan: In the indefinite construction that means the same as named?
Dan: Abrams is a Browne --- I assume that's doing a lot of the same work as in Abrams is Browne.
Johan: Maybe Mary is Browne, even though she's not named Browne at all. It means she's typical of the family.
Alex K: Earlier I proposed that you can say that 'is-equal-to-Browne' is predicated of Abrams.
Emily: What we were getting to isn't 'is-equal-to-Browne' but 'Browne' as the predicate.
Dick: If you take the view that proper names are concepts, then you're taking the intersection of those two sets and saying they're non-empty.
Alex K: Why do you want to talk about concepts and not individuals?
Dick: I'm a collectivist.
Alex K: But then you get into these problems about single sets.
Dick: Cleo and I wrote a paper on cardinality restrictions on concepts.
Dan: Abrams are Brownes isn't the same thing --- it's something radically different.
Alex K: Because it's a subset.
Dick: It's a one-directional intersection.
Alex K: Going back from proper nouns to copulas...
oe: In the FGD analysis, the role assignment is reversed?
Silvie: It's up to the annotator, and the rule is that you show what can be named what?
Tim: Isn't the left to right information structure?
Silvie: The focus here is the theory. What we name the theory now is the fact that Browne arrived --- it has the property of being the theory that we are concerned about.
Dan: That Browne arrived was our theory would be happier?
Silvie: The annotators are already used to identifying/finding what the individual and what the category is. If this were full-fledged annotation with topic/focus annotation, then theory would be to the right.
Dan: Alex's exhaustion was that it's half past five
ALex K: Are there any examples that people are really interested in?
Johan: What about definitionals, like A whale is a mammal.
Alex: That's generic. You've got to have another relation there.
Dan: I do nothing interesting.
Tim: We do set-subset.
Johan: Right, it's directional. You can't do A mammal is a whale.
Dan: One mammal is a whale. is better.
Dick: *A manager is Browne./Browne is a manger.
Dan: One manger is Browne.
Johan: In apposition, do you have the same identity relation?
Dan: I call it something different. I call it appos_rel. What I could prefer to do is use the new creatures in MRS called individual constraints (ICONS). I'm inclined to say that in John, the manager, just bought a boat. it's an ICONS.
Alex: Careful with that. I like the fact that you've got a different predication because appositives are presupposed.
Dan: I'm going to keep them different, but I don't want the appos_rel between two noun relations, because of quantifier scope.
oe: Nevermind the scopal position, technically there's a family of identity or equality relations, which is a fair high-level summary.
Alex: But they have a different effect at the discourse level, so we don't want the grammar to wash that out.
Dick: What's really interesting from this discussion is that it's not an equality relation, but a directional subsetting relation.
Dan: Managers are idiots -- there's some kind of subset relation. These guys are included in those guys.
Dick: Even something like Browns is old. What's interesting about that is that if you take it as being subsetting rather than equality, that actually makes the adjective and PP complements look much more similar to the noun ones.
Dan: I agree with John is a manager --- that really ought to be a predication. But John is Browne still feels different. We don't want predication in that case? And likewise, The theory is... you want that to be predication?
Johan: No identity.
Dan: My favorite theory is that there's more than one kind of copula.
Emily: To Woodley: Do you want that to mean My theory is that there's more than one kind of copula and My theory is my favorite.
Dan: Favorite means one among of a set of theories, and if the proposition is already in that package, you can't then pick out the one theory that is associated with the proposition.
oe: We don't want that proposition in the RESTR of the quantifier. Isn't that what you mean?
Dan: No... Woodley wants to treat it like The picture is of Bill.
Woodley: Do you give it to me for that one?
Dan: No: My favorite picture is of Bill. I don't want it to be that way there.
Emily: Isn't this what oe was saying about not wanting that in restriction of the quantifier.
Woodley: I agree that might be a problem.
Alex: The only conviction was for burglary v. The only conviction for burglary are very different. You've got to set up in the discourse context an alternative set.
Emily: The only conviction reversed was for burglary/The only conviction for burglary was reversed.
Alex K: These are all definite descriptions --- these are things that happen in the presupposition, not the assertion, so Woodley doesn't have a problem with that.
Alex: Yes he does! What are you having the presupposition of. You've lost the trace on whether it's a copula or not. How am I going to know there was a presupposition trigger there or not.
Alex K: You need a noun shaped referent to hang the propositions on and to quantifier over. And you (Woodley) only want to give her a preposition-shaped referent.
Woodley: Yeah, my theory doesn't explain everything yet.
Alex K: You could look at it as like a reified proposition or like a proposition itself. I wonder if one is like an atomic thing that the grammar provides and the other comes from light-weight inferencing (in my head). Let's say that the thing that Dan likes is what the grammar gives, and yet very frequently it gets translated into the other thing--attributing disasterness to something. Would that be a kind of set up that would satisfy you? I feel like it would satisfy me, except that having the extra inference, it's not as lean.
Woodley: I'd like the proposition to show up inside the scope of what I consider the relational noun and ideally only in its scope. Maybe it's better to say The theory is the theory that Browne arrived.
Emily: What do you mean by "say"? You want the MRS to show that?
Alex: In terms of paraphrase, it would get you past the alternative set for presuppositions.