oe: … under certain assumptions about the object language.

Woodley: I agree with the goal of abstracting away from the object language, but to make decisions about the design MRS based on the comfort that yes it would be possible to assign some interpretation to that, rather than whether it would be an interpretation someone would want misses the point about why we want a meaning representation.

Dan: example?

Woodley: If you're going to try to assign an actual LF in some object language that looks like anything I've ever seen, it's going to look very different depending on whether the argument is scopal or just a variable and what you can do with that (model builder, theorem prover etc) … the shape of those predicates being different is going to be relevant. The information that they have the same name and are related is also of course relevant, so they don't have to be two totally separate things. Don't have to decide now what the relation between _believe_v_1 and _believe_v_2 (nouny v. verby arguments) is.

oe: What in the syntax is maybe a much softer boundary/comparable distribution, in logical forms are very different things: operators v. predicates taking logical variables or constants. The reason Ann and Alex have to hold strongly to the position that in a real MRS each predicate has fixed arity (n scopal and m non-scopal arguments) is because they have to maintain the possibility of a bridge into conventional predicate calculus.

oe: Underspecification is often motivated by synsem interface and parsing efficiency etc. But OTOH trying to understand better the other end (relation to logic) is also interesting. Ann and Alex are probably much better equipped in that space. Save for Ann time?

oe: Reading up on Montagovian and Davidsonian semantics and trying to understand the relation to what we're doing.

Woodley: Not necessarily incompatible, but there's a gap.

oe: A big gap, due in part to the range of phenomena we're trying to approach, and self-imposed problems that Montagovian semantics is trying to solve.

Francis: Under the influence of rabid functionalists who hold the position that LF is not the correct thing to do, worth keeping in mind that LF is a way of expressing things, but it's not a given that it's what humans are doing when they're interpreting language.

Emily: Regardless of how what we're doing relates to what people actually do, it's formal so it has to be internally consistent.

oe: But that's a different position from saying it has to support inference.

Woodley: If not, what do we want to do with it?

Woodley: That leads to a nice broad question: What's the point?

Dan: Tiny point: There is a move we can make in the grammars we write to exploit the naming conventions a little more like their original design, since we don't usually know much about the actual sense distinction. Maybe we should be saying this is _discover_v_rel, and I can't tell you whether it's the one relevant for an index or a proposition. You go look in the SEM-I and see which one it must be depending on whether it has a scopal or individual argument. Not currently using _discover_v_1_rel consistently in this way.

Francis: In a case where we there's two very different discovers --- a cross-classification between different types of arguments and other kinds of sense distinctions e.g. I see as in I physically see v. I understand. Want to keep that distinction while folding in I see what you're doing and I see your point.

Woodley: Can mentally see either a NP or proposition, but can only physically see a NP

Emily: I saw him swimming?

Dan: I can see that he's swimming.

oe: I came to the pool and saw that my daughter was swimming for the first time.

Emily/Francis: So there's two dimensions of underspecification: sense distinctions and types of arguments.

Francis: So in that case, still all the same predicate? (If we can find one where syntactic factors can at least sometimes tell us which sense we have.)

oe: Classic example in that space is the count/mass paper. Currently we're looking for something parallel in the verb space. It's almost an hour into the meeting, so time to bring up the causative/inchoative alternation. No doubt we want a predicate distinction between The window opened/Kim opened the window.

Emily: Can't decompose?

Dan: Can. Maybe think of it as a notational variant.

Dan: The windowed opened onto a beautiful vista. The glass broke. The earth shook.

Woodley: Say that The glass broke is one place, and Kim broke the glass is a different predicate with two arguments.

Dan: _break_v_1_rel and _break_v_cause_rel: different predicates, different roles for ARG1.

oe: We've felt that what we're doing is equally adequate to the decomposed representation.

Francis: With explicit links would be equivalent. But some things don't take the alternation and then you have to decide which one…

Francis: When we talk about predicates, for the discussion have a distinction between predicates that are completely different and those that share something:


Francis: I thought many years ago we agreed that predicates in the MRS would have a fixed arity. Though implementation may not have followed.

oe: May have to be more deliberate about what we mean by MRS. MRS: an underspecified description of a set of logical forms. Ann and Alex consistently say that they have fixed arity. Grammars find it convenient to break that, so the things that grammars produce are underspecified descriptions of MRSs.

Emily: Even if predicates have fixed arity, it's useful to keep the description between MRS-predicates and LF-relations.

oe: Three levels we can discuss: underspecified descriptions of MRSs (one we love and understand very well), MRSs (one that is formally defined) and the logic language (on that is least understood).


UDMRS (Underspecified descriptions of MRS) predicate (possibly variable arity)


Object language (logic) relations: fixed arity

Francis: Leave room for semantic root at the top.

Emily: I don't understand the dimension that lines up semantic root with these other things. I think that's a separate independent dimension.

oe: Orthogonal.

Emily: Do we need distinguished terms for UDMRS and MRS?

Woodley: Having the same terms for different levels leads to confusion.

Ann: Don't think you need to make the grammars more structured, can go down the RMRS route and get an underspecified version of that predicate symbol.

Woodley: What would the implications be of the same UDMRS predicate having different uses with scopal/non-scopal arguments?

Ann: Don't know. The Alex et al formalization of RMRS didn't address this.

Woodley: not claiming that a single UDMRS would resolve in one case to scopal and non-scopal arguments, but rather that two different UDMRS produce by a grammar using the same underspecified predicate name might resolve to those situations.

Ann: But that any MRS that came out, you would know which it was.

Emily: With the help of the SEM-I.

Ann: The MRS coming out of the grammar tells you which way it is deterministically. Does it? If not you have the underspecification problem where you don't have any clue what the tree is.

Woodley: Looking at the predicate symbol plus what's assigned to fill that role does tell you.

(Emily: What about dropped arguments?)

Ann: I don't see that there's any serious problem with that. The external MRS should probably specialize one way or the other.

Woodley: So that's the MRS that's described the UDMRS produced by the grammar.

Ann: That's the MRS that goes through some well-specified …

oe: That is a well-formed MRS where the same predicate symbol is always used with the same arity of scopal and non-scopal arguments.

Ann: With ordinary arguments (individual), I think you could have an external MRS which underspecified the arity and fitted in with the general RMRS way of doing things. If the grammar deterministically tells you the type of each argument, could straightforwardly figure out which predicate symbol to use.

oe: Unless some of these arguments were unexpressed.

oe: At Cambridge, didn't Alex say that even in RMRS you have to assume that predicates have a fixed arity. Even eat would have to be broken down into eat/1 and eat/2.

Ann: If it's going to be a predicate in logical terms yes, but the RMRS universe does allow that sort of underspecification to happen. That's the whole point of allowing the RMRS fluidity about number of arguments. The RMRS language allows you to not know ahead of time how many of those things you've got.

Emily: The grammar question is whether we get dropped arguments of the ambiguous type, like I see.

Dan: No, the grammar prohibits that.

Dan: Two separate entries for see with different types of arguments, that happen to use the same predicate symbol. But only the NP complement one can drop its argument.

oe: I thought the move to RMRS gave the additional great flexibility of underspecification of argument roles v. modifiers.

Ann: The move to RMRS gives us the ability to have underspecified role names, and role names which aren't mentioned at all in the grammar, and to do things like have just an ARG1 and an ARG3 mentioned in the grammar.

oe: What means 'mentioning' in the grammar? The grammar already does this with the current UDMRS.

Ann: The assumption is that if there is an ARG1 and an ARG3 there is some unexpressed ARG2. But the point here is that with a standard approach to semantics you can't leave out some argument, but with RMRS you can. Gives you the option of not mentioning all the arguments which in the fully resolved MRS you do have; hallucinate the rest based on the SEM-I. The formal way of doing that is fairly well understood.

Emily: What about under specifying the argument/modifier distinction?

oe: Currently make a black and white distinction between Ps that are semantically vacuous and just mark a role v. those that introduce their own predication.

Dan: the rejection of Sandy seems to be a place where we'd like to have an underspecification. Sandy could be ARG1, ARG2 and yet another where Sandy is the vague possessor of that rejection. We would paint that in the loosest case as just a poss_rel. Maybe that rejection of Sandy's.

Ann: How would you want the poss_rel to go.

Dan: We were thinking that the link between those indices is either an ARG1 link, or there's another predication in there.

Ann: Can't don't that in the RMRS universe (as underspecification) but could imagine someone else making another universe where you could do that. poss_rel is a completely different type of object in the RMRS universe from ARG1 or ARG2.

Francis: But if prepositions are collapsed into semantic roles, could underspecify over that larger set of role.

Ann: Could imagine an approach where you make the poss_rel a special sort of del like arg1_rel … but probably better/easier to go down the route that Francis is suggesting where you invented another set of role names. But that would make for a different LF for the possession of rejection by Sandy from what we have at the moment.

Woodley: Doesn't that somewhat close to the door to the possibility that Ps have modifiers of their own? (The letter apparently to Sandy)

Francis: Yes

Ann: If you think of this in terms of things that we do think of as case marking appositions, there are arguments that those shouldn't be treated as case marking. That possession relation is already very underspecified. Need to work out a way of making that underspecification (like its kin the compound relation) cache out as… normally thought of as relations between entities, as opposed to ARG1 which is thought of as a relation between an anchor and an entity.

oe: What forced the interpretation of anchors instead of events?

Ann: We never talked about it (ARG1) hanging on the event.

oe: But that's classic Parsons-style. I know you started with handles, and that forced you into in-groups…

Dan: Was it that at the time we hadn't settled on the notion of characteristic variables?

Ann: Yes that's true, but there's another problem, even if you have characteristic variables. In some sense I think of this as a bit irrelevant as I'm now in the DMRS universe where many of these distinctions have disappeared anyway. It really doesn't matter whether you say that the arc of an argument relation is going from an event or an anchor or even just from the predication entity itself as it is in the DMRS universe. The point is that poss_rel is still a different type of thing (e.g. a node in the DMRS universe). Could change that and turn it into an arc in the DMRS universe, but then you've changed the logic a lot . I've always had the assumption that we have rather a small number of possible links (DMRS) or argument names (RMRS) and the specialization can go down to ARG1/2/3 but that there's not some sort of indefinitely large number of different roles, even if you went to a sort of more FrameNet-style approach, I've always assumed there was a finite, small, enumerable set of those things. There's a fairly fundamental distinction between the relation entities and the link entities.

Woodley: Still it's possible to think about underspecifying between arg/adj for PPs… I put the book on the table

Ann: That ones a special case (resultative). But Kim ran to the store, some people would say that's a complement syntactically some say it's an adjunct complement, but semantically it's the same.

oe: Currently we say that's a syntactic complement that introduces modifier semantics.

Ann: modifier semantics is not a good term either, since modification is a syntactic notion.

Dan: Kim put the book on the table.: on builds a proposition. I don't think we have any cases where a preposition is two-place predication and is directly the argument of a verb.

I handed the book to Sandy: hand treated as a three place relation, like give.

What I think we don't have is any place where the preposition lands directly as one of the arguments.

Emily: Prepositions can:

(1) Show up as a separate proposition (2) Show up as predicated of the verb (3) Not be represented in the semantics.

oe: What's wrong with the semantic modifier terminology? Ann: intersective modification is ok.

Woodley: I thought we found some examples of complement/adjunct ambiguity over the summer.

Emily: I think those cases are all really attachment ambiguities, not role ambiguities.

Dan: We did look at some cases where there might be some benefit in ambiguity of which argument is being filled.

Kim was seen by Sandy.: don't know if that's locative or passive agent; treebankers have to pick (possibly arbitrarily)

Would be nice to be able to say we just don't know.

Emily: But that's not role underspecification.

Francis: A slightly different view is that we're possibly missing a generalization in something like I gave it to Sandy and I read it to Sandy: in both there's some sort of benefactive role for Sandy, but we're not capturing it.

Dan: We're not? What about I read Sandy a book. That's just a gap in the lexicon: I read Sandy a book

Ann: Would have to do that with a very productive lexical rule, and need to constrain that.

Dan: There should be a lexical rule that is very productive and surprisingly idiosyncratic in its

Emily: And you want that to be ARG3 in all cases?

Dan: give is the canonical instance. Told Sandy a story.

Emily: I ran Sandy a marathon?

Dan: Utterly different. One is canonical, occurs all the time. The other …

Francis: Seems strange to say these are ARG3s.

Dan: But even if they're not all ARG3s, then surely you would say that hand requires there arguments.

Francis: I think we are missing a generalization. But I don't have a solution.

oe: I can recommend a solution: We don't have capture all generalizations. Surely there are meaning postulates…

Emily: A little bit of semantics textbooks is a dangerous thing…

Dan: Maybe RMRS gives us the flexibility we want. Maybe there's a ben role (extend the range of roles a little bit). Lexicalized cases like hand still assigns the ben role, since ARG3 is heavy-handed.

Woodley: Slippery slope?

Dan: Reduces the distances. sing is always two place, but there's this other benefactive that can show up.

Emily: So it's a syntactic fact that hand is always three place, where sing is somethings two-place?

Dan: Yeah.

Ann: Other issues re whether you want to lose the to/for distinction in that putative benefactive. I read it to Sandy/I read it for Sandy. I'm actually fairly neutral about whether we have a benefactive role as opposed to ARG3, then you have to ask whether you are losing something that you want. In the case of read, I don't think you can say the to/for are the same.

Dan: I wrote Kim a letter is underspecified between the two.

Ann: But if you just have a benefactive role, when the to and for are overt, you lose information that you have in the sentence.

Dan: Slippery slope… if it's only to/for, maybe we're okay.

Ann: I had worked on benefactive things years ago.

Dan: There is a potential hope that the way a language does valence changing things in the syntax corresponds to a limited range of distinctions in the semantics. So maybe we can still hope to live in a universe with a limited finite set of role names organized in a hierarchy…

Emily: Even if we're hoping for a finite set of role names we can't hope to have the same interpretation of them across predicates.

Ann: That's why ARG1/2/3 (because Dowty)

Dan: Could extend that to/for and say that the actual interpretation of to/for depends on the predicate that it goes with.

Ann: That's only possible if you've got the to/for in the MRS and if you're going to do the dative-alternation style thing so that you've got a role name for the NP NP case then you end up with MRSs that look very different if you have the to or for as a semantic thing.

Dan: imagining ben. ben_1 (to), ben_2 (for) [ben_2 and ben_4] … get underspecified ben in the NP NP case, and ben_2 or ben_4 if the P is overt or if the verb tells you have it has to ben (bake Sandy a cake is ben_4, not ben_2).

Ann: Losing information about what for means by saying it's benefactive.

Dan: The preposition will (through magic) produce a ben_4.

Ann: But when does it do that? The book for Sandy: is that for, or ben_4? Going to push the problem of ambiguity somewhere else if you're going to do this. For clearer cases, think of deverbal nouns… by going away from the surface representation with the for, have to decide where to draw the line, which moves the lack of generalization there.

Dan: I baked a cake for Sandy for my mom. and even I baked Sandy a cake for my mom. and I baked my mom a cake for Sandy with the same meaning. (But never: I baked my mom Sandy a cake, unless she's called Sandy.) I see that it doesn't scale very well.

Woodley/Ann: They are both benefactive in that example, which is an argument against benefactive actually being a role.

TheAbbey/Chrysalis2014Arity (last edited 2014-02-19 22:22:14 by EmilyBender)

(The DELPH-IN infrastructure is hosted at the University of Oslo)