Revising the algebra

Here are some additional comments which may be useful for the algebra discussion:

Clarifications from presentation

Ambiguity of term `INDEX', `LTOP' etc

In DMRS, the index is a node in the DMRS graph, rather than a distinct instance variable. Since null semantics items have no nodes, they cannot have an index in that sense. However, the feature structure including the DMRS can have a feature INDEX, which could, for instance, link to a complement's index. What the DMRS treatment precludes, which is possible in MRS, is that the index is used as a sort of storage facility for properties such as tense, which can be accessed even if there are no syntactic arguments. This has consequences (predictions?) for the association of semantic properties with things like existential `it' - while it isn't completely ruled out, it would require an additional structure to be added.

Existing algebra is broken with intrinsic variable grammars

The algebra as originally defined says that the semantic head provides the slot and the hook (the index and ltop plus xarg or whatever). This worked for adjective-noun combination when we didn't use event variables on adjectives, because the hooks were equated. Once one does this with event variables on adjectives, the slot is on the adjective but the INDEX of the adjective-noun combination has to come from the noun, so we don't have a notion of semantic headedness any more. With scopal modification (adverbs like `probably'), the index comes from the modified verb and the ltop comes from the adverb, hence the hook isn't transmitted as a whole.

This could perhaps be fixed but points to a deeper issue: when we originally developed the algebra, we (at least me, and I think Alex) were influenced by previous approaches to want a notion of functor and argument, with the functor (aka semantic head) providing the slot and the hook. This is not tenable without some operation that does something drastic to the functor. Arguably this is analogous to some of the type-raising operations in conventional approaches. e.g., in lambda calculus, one can make an NP into a functor. But we always thought of this as a problem with those approaches, so this doesn't seem like a good path to go down.

Another issue is that the connection to the syntax is looser than I would like. The syntax and semantics cannot be completely isomorphic (again, this is a divergence from categorial grammars of various types) but we really would like the relationship to be regular with limited exceptions.

Modest proposal

Basics

  1. We think of the algebra in terms of graphs (DAGs), since this is anyway what we're dealing with natively, and since it allows us to see the relationship of DMRS and other forms of MRS more clearly. Formalising in terms of graphs is possible but not done here, because it doesn't seem to me to improve clarity.
  2. The TFS graphs have the following type of structure (abstractly):
    • [ SYN [ SYNROLE1 [1]
      • SYNROLE2 [2] ]
      • SEM < [ PRED

        • SEMROLE1 [1]

          SEMROLE [2] ] > ]

    where [1], [2] indicates reentrancy. In what follows, however, for the purposes of the algebra, I'm going to try writing the semantic graphs with additional syntactic annotations rather than keeping the graphs distinct. I expect this to go wrong, but it gives us the idea that syntax and semantics usually go together and we have to do something to disconnect them. (The hope is that what we do is less violent/unconstrained than with alternative approaches. )
  3. As in the presentation, I will treat construction semantics as though it were lexical in what follows.
  4. In terms of the semantics, we have the following components, much as in the original algebra:
    1. INDEX b. LTOP c. XARG d. EPs and closed links between them e. open links (= slots)
    The most basic principle is that nothing can happen to things in category d. other than via the other categories. This is guaranteed in the TFS implementation by never having paths into lists and by not proliferating features beyond LTOP, INDEX, XARG. BUT we may have to modify this (possibly depending on what we assume about the syntax-semantics interface).
  5. Hook is retained in the dmrscomp grammar but has limited utility, since INDEX and LTOP no longer always go together. This would also apply to a revised algebra using MRS. In the dmrscomp grammar, open links always select for LTOP or INDEX, rather than accessing HOOK as a whole.
  6. All composition is treated as binary. This potentially allows some wiggle room in the syntax-semantics interface in grammars where there are n-ary rules with n > 2. Specifically X -> X1 X2 X3 could be treated as X' -> X1 X2, X -> X' X3 or as X' -> X2 X3, X -> X1 X' This is as in the old algebra. Whether we want to take this to the point where we (for algebra purposes) analyse examples like make him angry' as make (him angry)' despite the ERG binary branching, I am not sure.

  7. INDEX always comes from the syntactic head (see comments at top of page about ambiguity in terminology in DMRS). LTOP normally comes from the syntactic head, but comes from the modifier in scopal modifier examples (`probably sleeps') in the dmrscomp grammar and I expect there are other cases where this will be necessary.
  8. Syntax chooses the constituents, says which one is the syntactic head and chooses an open arc. The arc is typed for LTOP or INDEX, hence can choose its target.

Examples

for clarity, XARG omitted except when it's needed. Similarly /= and /neq

This is all DMRS-like, but the abstract MRS TFS graph can be obtained by adding nodes. There may be things we could do with that that we can't do with the DMRS-like representation, but (I think) not vice versa.

I hope the intuition is clear, but the notation needs some thought.

I like bagels

like <-ARG1.SUBJ- like -ARG2.COMP1->

both arguments select for INDEX. In DMRS, both links are /neq. In MRS, the NP LTOP is unconnected.

Bagels are liked by me

like + passive <-ARG1.COMP1- like -ARG2.SUBJ->

in terms of the old algebra (Copestake 2007 paper) the slots are transformed by the passive operation. In this notation, the syntactic labels on the open arcs are changed.

Bagels, I like

like + topicalization <-ARG1.COMP1- like -ARG2.SLASH->

Again, the syntactic labels on the open arcs are changed.

Kim tried to go

go -ARG1.SUBJ.XARG ->

[1]<-ARG1.SUBJ.XARG- try -ARG2.COMP1/h-> <target> -XARG->[1]

Kim seemed to go (raising)

go -ARG1.SUBJ.XARG ->

[1] <-SUBJ- seem -ARG1.COMP1/h-> <target> -SUBJ->[1]

When a node with a syntax only specification comes out of a target, the idea is that it may unify with an open link on the target. When it comes out of the main EP, a syntax only specification implies there is no semantic argument.

Kim made him angry

If we take the option of reconfiguring the syntax (see above), there is nothing unusual here. him angry' has the same semantics as he is angry' and `make' takes it as a scopal argument. If we don't do that, we may need to constrain this sort of pattern somehow.

<-ARG1.SUBJ- make -ARG2.COMP2-> <target> -XARG-> [1]

easy editor to try to impress

rather a guess here, since I haven't looked at its actual analysis in the ERG in any detail

to try to impress

[ 1 ] <-ARG1- try -ARG2/H-> impress -ARG1->[1]

easy editor (INDEX is editor)

easy -MOD/=-> editor

This illustrates two things. Firstly, this is an example where we need to retain `open' links in an English DMRS - we'd already decided to do this for ICONS purposes but here we need it in the same way as some of the other grammars. Secondly, we have two open links being simultaneously instantiated. This wasn't allowed for in the original algebra. Assuming we want to allow this, we need to work out how to constrain it.

Notes

[Scribe: EmilyBender]

Ann: The notion of the feature called INDEX is distinct from what that feature is pointing to. A DMRS has an index which is a node --- that's different from having a feature that points to that node. Semantically null things can have the feature INDEX, but they don't have any nodes --- that feature can't point to anything inside that.

Emily: What about dropped arguments?

Ann: We now believe we should add them to the DMRS in some sense. That's different. Expletive 'it' doesn't have semantic properties does it? The agreement things on the existing it aren't really semantic because there's nothing for it to be semantic about. Sort of a minor point; interesting difference between DMRS and MRS structures. Dropped arguments we can come back to. Ex. we drank implying we drank something --- I think I want those things to hang around for anaphora resolution. Also it seems necessary for the easy-adjective cases. It comes up in some other language… Japanese, right?

Mike: Dropped arguments shared between two different verbs.

Emily: Dropped arguments with agreement information.

Ann: Right---want those to stick around. Dropped arguments are semantics w/o syntax; expletive it exactly the other way around.

Mike: Recalling talk about a trigger rule to add them in in conversion.

Ann: I'm interested in a grammar generating DMRSs right at the start; for that we'd need those cases for Japanese, and for the easy adjectives. In terms of conversion from MRS to DMRS, we thought we'd have two versions. One with the dropped arguments and one without, because some applications don't care.

Guy: In cases when a verb has lots of different case frames, what do we consider the dropped arguments. Everything that you could have? know that v. know a thing.

Ann: I think that's taking us onto something that's off track for today. I tend want a frame for know that is know some proposition. The fact that that can turn up as a sentence-y type thing or as an NP is …

Guy: know a person ?

Ann: Different sense. know a fact v. know that someone won the race -- would try and make those the same sense.

Emily: That's a tangent --- we now that there's dropped arguments. It's a linguistic question, for our purposes today enough to know that they're needed somewhere.

Ann: Existing algebra is broken with intrinsic variable grammars --- realized it's a deeper problem than I originally thought. Notion of HOOK as INDEX + LTOP + XARG. In the algebra, there's always a slot that takes the hook as an argument --- equate values in slot with values in hook. Having the slot taken as property of semantic head, and the semantic head contributes the ltop & index to the result. This is broken now because e.g. adj+noun combination. Index has to come from the noun because the ARG0 is its own eventuality.

Emily: And its INDEX points to its own ARG0, which is needed.

Ann: Yes --- point of having the ARG0 in the first place is the point, for adjective modification. Else why would you want intrinsic variables. I think there's a better analysis of adjective modification…

Emily: I'm curious what that better analysis is.

Ann: Predicate modifiers: very(blue)

Emily: What about tense in predicative adjectives? Need INDEX = ARG0 then?

Ann: Even there, tense should really be something like past(blue).

Guy: What about interestingly blue -- interesting(blue)?

Ann: Not sure. But I think we've gotten too a point far down the slippery slope, with es all over the place. Would be nice to back up a bit. Our representation should allow someone to generate either the intrinsic variable semantics, or the more traditional form, or something in between. Not convinced that we need this semantics with all these eventualities in it. Not saying it's wrong, just that it's very unusual and we haven't fully justified it. What we originally thought about MRS was that it was a way of specifying different forms of semantics. Could do that if we allow ourselves the predicate modification.

Ann: Existing algebra is broken because we no longer have the notion of semantic head. We've got something like headedness, but one is where the index is come from and the other is which thing is supplying the argument, and they don't agree in the case of probably. Kim probably slept the index really has to be the event of sleep and the ltop has to be from probably.

Emily: What's the info that shows that the index has to come from sleep. Worry about this often with negation.

Ann: He did not sleep --- tense coming from auxiliary needs to get to slept. Also Probably slept deeply, if you decide that deeply attaches low, not an issue.

Emily: Kim did not sleep not convincing because did is picking up two complements.

Ann: Kim has probably slept.

Ann: Point is that the hook is not being passed along as whole; the ltop comes from probably, but the index doesn't. Notion of semantic head has been split. That's quite deep because we trying to think of this as functor-argument combination. If we do it this way, then that's not so clearly tenable. There isn't a really clean notion of semantic head anymore. The fact that the intrinsic argument property is there in the grammar breaks that fairly nice idea of a semantic head in the algebra, but it was always a bit squirrley anyway to say that probably is the semantic head but shares its index with the event of sleeping. Seems pretty basic to say that the index is associated with the syntactic head. So I'm not so sure I'm sad about that going.

Ann: Another thing was when I tried to work out the connection with the syntax rules, when worked out, was sort of loose. Not terribly well constrained. Some nice things about the way it worked. I managed to get the same approach to the algebra working with a non-lexical grammar like RASP, but it still seems looser than one would like. An idea from Categorial Grammar that syntax and semantics have to be isomorphic --- leads to all kinds of contortions to make it work and doesn't scale. Ex: the post-processing that Johann Bos does to get to semantics from the Clark & Curran grammar breaks what we would consider compositionality.

Emily: Wasn't Johann one of the ones in Berlin saying compositionality was overrated?

Ann: [Argument from learnability] He must believe in compositionality at some level.

Emily: I remember him articulating a position that the only thing that counts as compositional is strict functor-argument application.

Ann: We want syntax/semantics isomorphic in some sense, while allowing for exceptions where we have to. Want something fairly strong --- not because it will always work, but because the cases where it doesn't work are interesting.

Ann: The fact that that is broken lead me to think of this in terms of graphs, rather than functor/argument. That means don't need to worry about MRS v. DMRS. Once you convert MRS grammar to DMRS grammar, doesn't look that different. Structurally fairly similar. Worth going through this to see if we can come up with something that even if not technically an algebra, gives us some constraints. And I know Emily at least wants something implementable.

Emily: Yes, I'd be interested in trying to port this into the Matrix core…

Ann: Not clear yet if it can be implemented (cleanly), but that would be interesting. Let's see if the thing works on paper first.

Ann: When we have TFS graphs, we have a separate piece for the syntax & for the semantics, and then we have a linking theory. From the POV of the algebra, will just be writing semantic graphs with syntactic annotations. Looks (D)MRS like + syntactic names to the links. This will probably go wrong, but seems like a reasonable start. [continues reading assumptions from notes]

Ann: Most basic assumption is that only access EPs through LTOP, INDEX, XARG.

Emily: Does that include derivational morphology?

Ann: No, and I'm ignoring derivational morphology for now. I think of it as something that you can think of as either a way of capturing a partial constraint or as something that's fully productive. If it's a partial constraint, it can be erratic. Just the connection between the words and the semantics can be somewhat erratic (semantically empty words, cx semantics).

Emily: But completely productive derivational morphology should be as well behaved as the syntax.

Guy: If we use my system from HPSG last year, it's just the same as the syntax…

Ann: In dmrscomp grammar, kept HOOK, mostly for symmetry with mrscomp grammar, not accessed as a whole.

Ann: The previous algebra always treated composition as binary, because always talking about functor-argument. Ternary rule in the syntax binarizable two different ways in the semantics, might be useful. But not sure I want to keep binary-only.

Ann: INDEX always comes from the syntactic head, LTOP normally does (couple of cases where it doesn't in the DMRS grammar). NPs have the same thing as before in not having LTOP linked to their components. When you're trying to formalize what's going on in composition, the syntax chooses which arguments to combine, chooses the syntactic head, and chooses an open arc. Then the rest follows from what's set up in the semantics + constraints on the semantics.

Ann: Going through the examples on the page. The links have syntactic info embedding to abstract away from particular grammar. Not going to talk about LTOP except where it becomes relevant.

Emily: How do we know which of INDEX and LTOP is pointed at by ARG1.SUBJ? That's the type of the arc somehow?

Ann: It works out in the grammar that every time you have an arc like that, you know whether it's an INDEX or an LTOP. /eq and /neq always go for an instance variable, whereas /h and /heq always go for the LTOP. The /eq is always relating an LTOP. If you have a link which is /eq it has to be primarily going for the INDEX. The LTOP equation happens alongside. If you go through what's happening in DMRS if the target is an NP, it's always /neq going for the INDEX. If the thing is a scopal link (qeq -type, /h) then it's always going for the LTOP. The other cases are always /eq, which in some sense is going for both INDEX and LTOP, but the INDEX is primary.

Emily: Because the INDEX is what will be the argument, the LTOP there is just for identifying labels.

Ann: Yes, and you also get the case I was showing the other day that shows that you need them both.

Guy: Coordination?

Ann: Having both L-INDEX and L-HANDLE is wrong.

Emily: Yes.

Guy: Ignoring that case, for regular adj+noun, we need both INDEX and LTOP?

Ann: INDEX and LTOP will be the same thing in DMRS in ordinary MRS they're going to be different types of things, but in some sense the adjective is always going for the index.

Guy: What about Kim probably didn't leave.?

Emily: Only needs the LTOP.

Guy: Can you get non-scopal modifier outside scopal one.

Emily: Maybe we're predicting that that doesn't happen?

Ann: Seems not to come up

Mike: Kim quickly didn't leave.

Emily: Kim emphatically did not leave.

Ann: I think that's metalinguistic.

Mike: On L-INDEX/L-HANDLE: Kim probably slept and certainly ate.

Emily: Don't need both. Verby coordination points to handles, nouny to indices. (LTOP of nouns always left unlinked.)

Ann: Yes, it's a holdover from when Dan wanted one form of coordination for both, and then kept linking the indices for verby coordination because the could.

Guy: So /h LTOP, /eq and /neq INDEX?

Emily: And /eq also LTOP, because in non-D MRS, need to do label identification.

Guy: So do we really need separate nodes? What's the situation where they are two different nodes and you don't know what might combine next?

Ann: The weird case with the relative clause… maybe? I don't know, but I'm not prepared that you can never get non-scopal outside of scopal, so it might be that. We could look at that and see whether they are such cases.

Guy: So that's an open question then.

Ann: I think you probably need both. And I'm trying to keep in MRS-land as well as DMRS-land.

Ann: Back to examples --- can think of passive as swapping arguments around (it's more complicated, but still). Syntactic labels on the open arcs being changed in order to allow for a different syntactic operation like the normal way of talking about it in HPSG (for passive, topicalization).

Ann: More interesting examples Kim tried to go.

Emily: SUBJ.XARG is surprising to me.

Ann: It's not a path, it's two different labels for the same arc (not a path).

Emily: "<target>"?

Ann: Just a representation for a node.

Guy: Is XARG a node?

Ann: It's always a node that is from an open arc. In some sense it's more relevant to think of it as a label on an arc, than as a node.

Emily: Do we need a constraint that says that XARG always points to something else in the lexical entry, or else nothing external needs to grab it? (A grammar that doesn't obey this constraint would give broken MRSes…)

Ann: When you start in a lexical item, the XARG is always another label for an existing arc, but in the graph language, you sometimes label arcs with only XARG, but only when they're pointing to some other structure (not within the lexical item).

Guy: So the XARG is always a property of the node being combined with. Can't have more than one arc labeled XARG.

Ann: I'm assuming the feature structure convention where there's only one thing per label, but you're right that I'm only sort of half following that here.

Guy: There's two XARGs here, but there's only one coming out from any particular node.

Emily: Raising example. Do you want to do raising in terms of SUBJ of the complement not in terms of XARG?

Ann: In terms of the grammar at the moment, it shares the entire subject. Could be in terms of XARG or in terms of SUBJ. There's a reason syntactically to do it as SUBJ.

Emily: We do that in SWB, but I think it's wrong. We say that the whole synsem is the smallest thing that gets both the index and the syntactic info we need -- but that doesn't mean you need to grab the whole thing. The empirical issues have to do with expletive selection (moot in the ERG, because Dan uses the types under expo-index) and quirky case in Icelandic. But even those don't require grabbing the whole synsem. We can say that both raising and control predicates grab the XARG and that raising predicates additionally (in some languages) reference some syntactic feature(s).

Ann: Okay, good. XARG there is better.

Ann: Kim made him angry. If it's binary, no semantic link between make and him. Make has two complements, but the first one it doesn't link to semantically at all. him is the COMP1, not semantic link there.

Emily: What do you mean by reconfiguring the syntax?

Ann: In some sense ternary, but in the ERG, we treat everything as binary, and so naturally combine make with him, then that with angry. If we pretended that him angry was the constituent in the syntax, could combine those directly and make could take that as a scopal argument. Not saying we actually do that, that's one way of seeing what's going on in this example, which would mean that we don't need to … the point is that there's a certain amount of arbitrariness in the binary branching in the ERG. If you're going to insist on the thing being binary branching, but I actually make it go the other way.

Emily: Is it problematic in this version of the algebra, because that first combination has too many open pieces?

Ann: My feeling as of an hour ago is that I'd rather treat this as ternary and not say that things have to be binary. But it's simpler to treat things as binary, though, in terms of writing constraints.

Guy: One of the open arcs is pointing to what the target's XARG points to.

Ann: But COMP1 isn't really an open arc in terms of the semantics, because it doesn't have a semantic label.

Guy: I was asking a different question --- I still feel uncomfortable with XARG in these graphs. There's a single XARG node from an open arc. Because then we don't ever have to refer to links inside the target graph. There's an INDEX, LTOP, XARG and those are things that we can point to.

Emily: Yes, exactly, that's why XARG was in the HOOK.

Guy: Can we instead say that we pick up the target of the only open arc of the target?

Guy: When we do COMP2, we take the target graph and we associate the target graph's LTOP with make's ARG2 and the target graphs XARG with make's COMP1.

Ann: The target graphs XARG is the end of an open link.

Emily: If we're doing it binary, what does the first step look like?

Ann: The way the ERG does it? It doesn't work. Well, yes it works, but you have to postulate a hypothetical node as I was starting to talk about in the L-to-R example a couple of days ago. You have to invent a placeholder and then it can work.

Guy: Why doesn't it work?

Emily: The pattern is the lexical entry for make --- when we pick up him, it's only COMP1.

Guy: So can't it be COMP1?

Emily/Ann: That's only syntactic, not semantic.

Ann: It's a disconnected graph in the semantics, and that's something I was trying to avoid.

Joshua: But the other binarization works fine.

Emily: But it's terrible in the syntax.

Ann: I wouldn't try to do it in the syntax.

Guy: What's wrong with having a disconnected graph?

Ann: Disconnected graphs are what you try to avoid like the plague.

Guy: But it won't be disconnected in the end.

Ann: I think you want to try and avoid that. If you allow that, then you can basically squirrel away all sorts of stuff. Not very constrained once you allow that.

Guy: Still constraining how you can combine the graphs.

Ann: Only to the extent that lexical entries are constrained, which is not very. Not the definitive answer, just pointing out something I'm uncomfortable.

Guy: Why are disconnected graphs an issue? This structure is not complete.

Ann: I'm trying to make the pieces all connected graphs, but it seems like a nice thing to try to achieve.

Emily: And if we can't keep to that, it would be nice to find a suitably backed-off constraint, instead of anything goes.

Ann: Alternative is placeholder node that we haven't seen yet. The reason the example is in here, is because I've been worrying about.

Ann: The next one is these easy things.

Emily: Piece of cake!

Ann: Those are just horrible. Wouldn't have worked in the original algebra (and I've looked at make as well). An easy editor to try to impress. Filling slots in both directions, clearly not algebra compliant. Need to keeps open links in the sense that they're never bound.

Emily: How so? ARG1 of try/impress.

Ann: yes.

Emily: Why try to impress and not just impress?

Ann: Rules out some cheating ways of doing this.

Emily: So the problem here is the mutual selection which isn't just functor/argument. And we'd like to do this but can't in the old algebra?

Ann: We may like to, but how do we constrain it?

Emily: Implementation status? Something I can use to port to matrix.tdl?

Ann: dmrscomp grammar doesn't have any obvious ways of constraining what you're doing, just an attempt to get out the same sort of structures. Wanted to see if you really didn't need as separate INDEX node.

Guy: Is the XARG the subject of the target?

Emily: Oh right -- you want the SLASH of the target, not the XARG. We got rid of SUBJ before, but still have SLASH.

Ann: Trying to go from a semantic target to a piece of syntax for that target. Wanted to just do syntax at the top level, and not go in any further, just XARG.

Emily: To show this problem, and not the mutual selection thing, with This is an easy editor to try to impress. Don't we also see this problem with other LDDs?

Ann: I don't think so.

Emily: Bagels we think Kim likes.

Ann: we think Kim likes still has a SLASH.

Emily: So does to try to impress. How is it different?

Ann: It's because of easy.

Emily: So it's the fact that it's lexically mediated, and not a construction doing it?

Ann: Will attempt to make that a bit clearer.

Guy: Syntax picks out the SLASH and the semantics can just do its thing.

Ann: I wasn't really thinking SLASH --- more worried about about the mutual selection.

Guy: Bagels we think Kim likes has no mutual selection. Don't have to worry about the SLASH because the syntax is doing it. But easy we have something more complicated…

Ann: The mutual selection would be a problem for me, even if there wasn't a SLASH involved.

StanfordAlgebraAdditions (last edited 2016-06-21 13:47:08 by EmilyBender)

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