Skip to content

DiderotUnderspecifiedAttachment

GuyEmerson edited this page Jun 27, 2018 · 2 revisions

Berthold:

hat          im     Labor er es blitzen   sehen
have.3sg.RPS in.the lab   he it flash.INF see.INF

He has seen lightning in the lab
 -- lightning in the lab
 -- seeing in the lab

In lab and flash aren't a constituent; there was no way of underspecifying it. Idea instead (with ACONS) -- gather up targets and then resolve attachment later. Current GG solution: Just an arbitary attachment.

Ann: ACONS was a two-arg thing, with an index and then a list of targets, with the semantics of this index is coindexed with exactly one of that.

Dan: How do you take things off of that list? Because there are targets that are excluded.

Berthold: It must have been a list of pairings...

Emily: [Reviews big picture overview of Hankø notes, ending with:] came up with examples where the structure puts in some constraints. Is that what you meant Dan about taking things off the list?

Dan: Yes --- if the inner one attaches high, the next one can't attach lower.

Guy: Is that true with these German examples?

Berthold:

hat im     Labor [vom      Garten aus] er es blitzen sehen
has in.the lab   [from.the garden out] he it flash   see
'From the garden, he saw it flash in the lab'

Each modifier can still attach to either verb, high attachment of one doesn't block low attachment of the other.

Guy: What about "erblicken" --- is that the same kind of verb? With "sehen", we can't tell if it's an infinitive or a reduced participle (without "ge-").

Berthold: No, it's a very restricted class. But it includes "hören" -- and it's "hören", not "hört". Called an "Ersatzinfinitiv" (replacement infinitive).

hat im     Labor er es klingen  hören
has in.the lab   he it ring.INF hear.INF

Dan: On Tuesday I thought John was going to win. --- we currently need two different slash paths, can't underspecify. Does the same ambiguity happen in German with fronted PPs.

Berthold: Sure.

Dan: And that's still independent of verb clusters, mittelfeld, etc.

Berthold: The explanatory account has a trace down by the lower verb, and you move things around in the mittelfeld. That's nasty because you get two simultaneous movements. Can also get remnant movement: Blitzen sehen hat er es im Labor

Guy: To follow up on Dan's example, if the base version

Dan: When did you find out that Bill was going to the party?

Guy: But looking for an example where it's in the mittelfeld...

Dan: I'm looking for a way to use the same underspecification in the fronted case.

Guy: I wanted to see whether that was a different phenomenon from this German.

Dan: Could find different solutions, but if we can do underspecification, it would do both.

Ann: ACONS is indeed a list of pairs -- a list of (index, label) pairs. Can't reconstruct why.

Berthold: If you add negation ... or even here. Seeing qeqs flashing.

Ann: What are the ACONS you'd have for the flashing example?

Berthold: sehen: [e1, l1]; blitzen [e2, l2]; im Labor [e3, l3]

Ann: Here's what you sent me as a spec

isect-mod :=  avm &
    [mod-anc index-lbl-pair,
    target-ancs diff-list].

index-lbl-pair := index-type & lbl-type.

index-type := avm & [index individual].
lbl-type := avm & [lbl handle].

hook := index-type & [LTOP handle,
    XARG individual].

relation := lbl-type &
    [WLINK list,
    PRED predsort ].

Berthold/Ann: What's on ACONS is a list of isect-mod objects.

Berthold: Grammar builds up the target-ancs lists, and then the modifiers grab hold of them to add to ACONS.

Dan: So you're collecting up modifier targets before you know what the modifier can modifier. If you find an adverb that can only modify the verby things, but it gets the whole list.

Berthold: But you have types.

Guy: To ground this in data... what if it were an argument of the lower verb instead of in Labor.

hat  den        Kindern      er es helfen   sehen
have the.PL.DAT child.PL.DAT he it help.INF see.INF
`He saw it help the children.'

so there's a very strong constraint that it has to be associated with the lower verb...

Emily: But we aren't talking about doing this for complements.

Ann: Where are you putting the constraints, it has to be in terms of unification because that's all we've got. Are you putting it on the event?

Emily: In the case of a complement, the constraint isn't on the event. It's because the one verb is looking for a complement, and the NP itself can't modify anything. I think what Ann was saying is that in your adverb case, it can get that whole list of targets, and only the events are possible resolutions.

Dan: But if you have idiosyncratic syntactic constraints or something about configurations...

All: two-stage was actually about human disambiguation in treebanking, and secondarily in trained models. (Not about how we're modeling language...)

Dan: Brings up extraposed relative clause--A man saw a woman yesterday who I never met. Allowing underspecification would help here.

Berthold: In German that's disambiguated by gender.

Dan/Ann: ... which is on the index, so we can have the info needed in TARGET.

Ann: I think it's worth treating extraposed relative clauses mechanically as anaphora, even if we don't think that's really what's going on.

Dan: Or with a long-distance dependency, but that's really expensive.

Berthold: SOV German has much more relative clause extraposition.

Ann: Syntactically they're sitting somewhere, and potentially you have some sort of accumulator constraints, and that can't be done in ICONS, because it looks the same.

Dan: The worry is if one finds syntactically controlled island constraints where certain indices become inaccessible. Then we'd need some way of passing those cosntraints into the resolution engine.

Berthold: Like a right roof constraint.

Dan: Accumulate indices within a clause, but the constraint bearing construction would discard info from below.

Ann: But we're already not doing a good job of accumulating constraints on pronouns.

Berthold: ICONS wouldn't cut it because ICONS is only about pairs of individuals...

Dan: I think ICONS could work for pronouns, like Principles A&B.

Ann: Not those, the ones with quantifiers.

Emily: Do those ones require syntactic info somehow, or is it all visible in the structure of the MRS?

Ann: Potentially visible in the structure of the MRS. Might be some cases where we're not retaining enough information. The data is totally messy anyway. I had 10 marbles and I could see 9 of them on the floor. #It was under the sofa. You get to places where the constraints are unboundedly difficult to work out. (Not a syntactic constraint in this case.) If you construct a proper MRS model, you don't have a distinction between the 9 mentioned and the one not.

Emily: But if you're looking at the MRS you do.

Ann: It also interacts with presuppositions and stuff like that.

Emily: This thing about index,label pairs is easier in DMRS, right? Because then the modifiers could just say ARG1/EQ to that node.

Ann: In my email to Berthold, I was thinking of those things as full hooks.

Ann: In DMRS you also have potentially easier access than you do if you're just passing the HOOK around. If just passing the HOOK around you don't have a way of going back up and seeing other stuff.

Emily: So how far back along the graph can you go?

Ann: Depends on the type of operation. If you're looking it as feature structures you can't go back; if it's an operation on MRS structures qua MRS structures, in e.g. a separate resolution stage which is what I implemented for Berthold.

Emily: I wonder if that would give enough leeway for the examples where the high attachment of one blocks low attachment of the other. But probably not, because for that you'd need the order which is definitely a syntactic fact.

Ann: If you do allow yourself access to the order, then there's a nice solution in terms of no crossing dependencies.

Dan: And we do have the characterization.

Guy: What about an 'output' list for each attachment.

Ann: But the point is that there's an elegant solution in terms of dependencies that can't be reproduced without the order.

Emily: And order of the modifiers.

Guy: Put the marbles in the box on the table.

Ann: So not the more twisty ones like Put the apple in the fridge in the box. --- with the right intonation, works with there's a box in the fridge and you have to put the apple in the box.

Guy: Can't both modify put?

Dan: No --- because it's not a modifier, it's a complement and there's only one slot.

Ann: There's an argument going back to the 1980s that in the fridge in the box is providing one argument with zero conjunction.

Dan: I do that in the ERG now. pp-mod rule where they attach together and they're both waiting to serve some higher master. No conjunction relation because EPs with the same scope are conjoined anyway.

Emily: Do you do that for In Paris on Tuesday Kim and Sandy went for a walk?

Dan: Yes, there too. Used to have one modifying the other, but that's not what that means. Zero coodination rule cleans it up.

Guy: 2018 version of the ERG?

Dan: Yes. Paul Haley finally cornered me with some examples.

Berthold: Fridge example reminds me of:

Sollen wir in März  noch    einen Termin      machen?
shall  we  in March another one   appointment make?

Dan: If we were going to adopt this underspecification, would we do it across the board, including in I reserved a room for Sandy, where it's easy to build two trees, but we don't know how to choose. Also, how to force the grammar to only attach PPs at the highest point? Found it to be hard.

Ann: Why not lowest point?

Dan: The lowest point strategy is sort of viable. But then the accumulation looks a little different, because the set of targets isn't available at the point of attachment.

Berthold: Then using CFROM/CTO to geometrically constrain the attachments?

Dan: I was also exploring adding negative constraints in another place.

Emily: How can you do the blocking of low attachment of later modifiers based on earlier ones as something you accumulate as you go up the tree?

Dan: The later attaching modifier can see the earlier attaching one being there and add a note relative to it.

Ann: Then you need a Catalan series number of constraints.

Dan: Rarely looked at more than three...

Guy: Not enough to just say higher than that one?

Ann: If you just use CFROM/CTO can do it with the simple examples like PP sequence and compound nouns. The cases where it goes wrong is with the coordination possibility. The geometric structure thing although pretty is not enough but it doesn't allow for all of the combinations seen in the data. Not rich enough.

Guy: The question would be what's the scope of how much we want to underspecify. The geometric one also wouldn't allow underspecification of put in the fridge in the box (coordination case)---still have a two-way ambiguity there. If we want to underspecify over that was well.

Dan: Also that's a complementation issue--HCR v. HMR. We're going to keep that ambiguity anyway. The ambiguity we should be trying to reduce or eliminate is different ways of doing the HMR structures.

Guy: Just leaving the modifiers underspecified?

Emily: If so, then does the geometric solution come back into play.

Ann: I remember there being a strong reason not to pursue it.

Dan: Me too, but don't remember what it was.

Ann: Something about the interaction with the syntax.

Emily: Anything that requires the order to be passed into the semantics offends my sense of asthetics.

Dan: But in the case of noun noun compounds (without other stuff interspersed anyway), the syntax adds nothing beyond the order.

Ann: We already pass information into the MRS on indices that isn't purely semantic.

Emily: PNG?

Ann: You have to have plurality on indices, but it doesn't have a direct correlate in the model. But you need it (and gender) for pronoun resolution. You're attaching that information where you really don't want to say it's semantic. (Though Dowty has a paper saying that German grammatical gender is semantic. But I'd rather say it's not semantic, it's in the MRS.)

Emily: Have we reached any conclusions?

Dan: Done a lot of reminding ourselves of previous conversations and blockges. Don't have a strong sense of this being really viable---probably memory of previous attempt. It was hard and the efficiency cost was disastrous, because this kind of ambiguity packs really well/quickly.

Ann: If you have DMRSs and you pack DMRSs, it's really potentially quite elegant in terms of the way the structures collapse when you're doing incremental resolution. The packed resolution in the compound case is pretty straightforward. As you go along one of those bracketing and the others collapse. Master's thesis a few years ago --- code is checked in for DMRS packing. Not integrated with the grammar, taking DMRSs and putting them together.

Ann: So just do what you're doing.

Dan: Except it doesn't solve Berthold's problem.

Ann: ACONS does though.

Guy: And this one doesn't have the geometric problem.

Dan: Maybe this is just overreaching to see how broadly we can expand the ACONS solution ... but it's still potentially quite interesting for the problem that gave rise to it (German configuration).

Berthold: Could extend it to the put examples, if put exposes its complements as well in the TARGET list.

Emily: But put requires that complement position, can't just say --- here's a thing something can attach to.

Guy: Put it in the fridge in the box has one reading that Find it in the fridge in the box doesn't have, because of the zero coordination.

Emily: That's not allowed?

Dan: Right I don't allow the zero coordinated PPs to be modifiers. Because that would mean just the same thing as each modifier attaching independently.

Emily: But you do allow them to be extracted modifiers?

Dan: Yes. Engineering hack to allow them there while blocking them in the other place.

Ann: Coming back to the packing situation, just because I don't know whether there's some scope for talking about this in terms of the interaction with the machinery, maybe long term. The version with the DMRS is that you have a series of graph fragments and equality constraints marked in terms of the parses that contain them. Each node is represented only once, but associated with potentially hundreds of parse numbers. Similarly with equality relationships between ARG links and what they connect to. Big disjunction labeled in terms of the parses in which its true. As you find that certain equalities are true/false can collapse the search space as in treebanking. If I talk to Woodley very nicely, and if we had grammars that were producing DMRSs directly, could we set things so that the packed forest representation that he produces could be converted to the packed DMRS representation? Would potentially be an object of interest to system builders.

Dan: A smaller step: Could we import the DMRS forest pruning mechanism into the unpacking routine? Could use it to thin the forest before doing any parse ranking. First step would be to throw away everything that doesn't satisfy the DMRS constraint.

Ann: I see this as an interface between what Woodley's got and what you'd want to give to an outside system. Say you're talking about a microworld, have a constraint where you'd want to avoid all analyses that don't include that constraint.

Dan: Would be nice to do it as I unpack.

Ann: I was wondering whether you could take the packed feature structures and run a DMRS conversion machinery over the packed object.

Dan: He does some miraculous things without unpacking.

Ann: Don't need to manipulate, presumably it's just a change of formalism.

Emily/Dan: But the MRS isn't used in packing/is on the restrictor/would block things from packing if included.

Dan/Ann: Empirical questions: (1) Let's keep DMRSs in our structures as we pack them will pack less efficiently--how much so? Then when we start unpacking if we have a DMRS constraint from the target domain. (2) Can we translate from Woodley's packed representation directly into a packed DMRS representation that is goig to be easier to work with.

Ann: It may already be that Woodley's packing would work as is over DMRS feature structures.

Dan: Another path: Let's pack as we usually do and then find some mapping from that packed forest into a packed DMRS representation. Would be ideal if we could do it without full unpacking. Then we could experiment on what we could do with a packed DMRS. And then we could come back to the engineering staff and ask whether we can get DMRS packing efficiently in one step.

Ann: What's wanted is a way to ask the downstream application for input on key points to collapse the ambiguity space.

Dan: Might even improve speed not just accuracy of unpacking. That's a big cost at least in ACE.

Ann: Email Woodley a toy grammar and constructed very ambiguous sentences that the grammar can handle. What happens when we don't put the semantics on the restrictor---what does the packed representation look like.

Dan: One version of the experiment should be doable without any change to ACE. If we just put the DMRS in as we pack we can see if it plays well with packing.

Ann: Are there any tools for looking at the packed forest? FFTB?

Dan: He's computing the disciminants without unpacking...

Clone this wiki locally