Woodley: What’s the point?
- Is this a string of the language, but more than that.
- Emily: Who did what to whom?
- Francis: Do I know enough about this to say it in another language?
oe: Then we might as well stop…
Woodley: Questions like which of these representations are better, could make sense to view that from the perspective of better for what?
oe: That’s utilitarian.
Woodley: Did you have a different type of science in mind?
Emily: If you’re not asking better for what, that assumes you have some sort of ground truth to compare it to.
Woodley: There’s intrinsic or extrinsic, and for intrinsic, what do you compare it to?
Woodley: The argument/adjunct distinction discussion was interesting, because I was convinced there was something we were failing to underspecify that could be grasped by RMRS, but that seems not to exist. But if it did exist, how would we tell which is better?
Woodley: Dan’s question of why distinguish between _dog_n_rel and named_rel [ CARG Fido ], but how do you approach the question of why these are better?
Woodley: The design principles could be a stand-in for intrinsic evaluation of design choices. But in cases where that doesn’t fully determine there should be some way of still having progress.
oe: Close paraphrases should lead to similar or identical structure. I think we have intuitions about what is a close paraphrase and about how well we achieve that.
Woodley: The window opened/The window was opened
oe: Kim gave a book to Sandy/Kim gave Sandy a book; The fierce dog/The dog is fierce
Francis: I went yesterday/Yesterday I went still info-str differences.
Woodley: We have a list of these close paraphrases where that principle has already been applied and makes great sense. We’ve been bumping up against cases where that seem to quite get there, because our intuition doesn’t go that far.
oe: Can be that, but there can be cases of competing forces/design principles in competition.
Dan: In quite a few cases there are two different near targets which diverge.
Woodley: Non transitive property of being similar.
Francis: I lost my temper/I lost my cool/I got angry
Dan: Mostly stayed pretty close to the metal in terms of close paraphrases having the same vocabulary. As soon as there’s a different content word, don’t try to get the paraphrases Kim’s age is 10 years/Kim is 10 years old not trying to make these the same.
Woodley: Certainly would be nice to have something to say about that relationship, but not necessarily right out of the grammar.
Emily: Meaning postulates!
Dan: Meaning postulates are a very powerful tool, why not use it for Kim gave a book to Sandy/Kim gave Sandy a book.
oe: Notion of economy.
Woodley: If the meaning postulates aren’t written down anywhere, not very satisfied.
Dan: Now have about 100 meaning postulates written down (implemented) for logic course. Not all actually meaning preserving. A and B are cubes/A and B are both cubes/Both A and B are cubes (1st and 3rd already have same semantics) also A B and C are all cubes/A B and C are cubes. David B-P insists they are exactly paraphrases.
Emily: This is where I think the logicians are interested in the full range of meaning of NL.
Dan: No, but they have the advantage that they can be precise, and they have a good test.
Emily: So they’ve carve out a small space. Most NLP applications aren’t doing block worlds.
Dan: Even that small space is hellishly hard. I haven’t seen any cases where the NLP application wants something that breaks what the block world wants — more like sentiment analysis just doesn’t care about those things.
Emily: Don’t you have the linguist’s intuition that A and B are cubes means something different from Both A and B are cubes?
Dan: No.
Emily/Woodley: Maybe truth conditionally equivalent but would be said in different situations/different distributions.
Francis: Would be interested in trying slightly larger block-world type applications
Dan: I bring it up because it is an interesting application. Can try lots of different linguistic things.
oe: One approach to meaning postulates is MRS rewriting (like what Dan does). Another would be to map into a suitable form of logic and use a reasoner.
Woodley: That one is out of reach because we don’t have a suitable object language.
oe: How much damage do I have to do to use something that can be used for reasoning. How much damage would I have to do use FOL? Would we lose any embedded clauses?
Emily: Yes, but Josh Cason says …
oe: Third approach currently exploring in Oslo is dumping EDS/bilexical dependencies/MRS to RDF graphs, in the first instance to make them searchable. Worked out how to do that and build an index that can be efficiently searched for DeepBank and WikiWoods. Then there are also reasoners. Would be interesting to ask how much damage will result from interpreting MRSs
Emily: It is a lossy transformation, but is the loss relevant for some task at hand? (Back to Woodley’s question.)
Woodley: Lossy; defining meaning postulates over the lossy representation. If the task was to make statements about full-featured meanings, not clear what you’ve gained.
oe: In MRS rewriting space can work with native representations, but don’t have a formal notion of inference. In logic, lose some expressivity, but have a well-defined notion of inference. Trade-off. Would be interested to see what we lose in expressivity. Make one attempt to reach out to an object language and understand what the limitations are that would be painful.
Woodley: Have to decide about quantifier scope.
Emily: Is that true about description logics? Aren’t they quantifier-free?
oe: I believe so, no room to represent quantifier scope.
Dan: most dogs bark is something you can’t express?
oe: Express is one thing, but in terms of inference…
Dan: Can I infer from most dogs bark that more than half the dogs bark?
Francis: In DL, no, I believe.
Woodley: Isn’t that the kind of meaning postulate you’d like to write?
Dan: Although >50% is not a good paraphrase of most for non-logicians.
Emily: If we map to a target object language, we can see what’s lost, but to experience ‘pain’, we have to have a task, else it’s just aesthetic pain.
oe: J. Bos gives the impression that he can do adequate NL semantics in FOL. I don’t expect that be to true, but can’t say exactly where I’d feel what he does is not adequate.
Woodley: I think you can describe the shape of an MRS in FOL, but what you interpret it to mean… end up with predicates for dog that have a possible world argument and a time argument; it’s going to get hairy, but probably possible.
Emily: Do you mean predicate logic with GQ, or do you really mean FOL?
Woodley: FOL.
Emily: And embedded clauses.
Woodley: I think they end up with some kind of possible world analysis.
Dan: John knows that Mary sings. What does John know in FOL representation?
Woodley: He knows a fact.
Dan: John denies that Mary sings. Then there is no fact; have to have a proposition.
Woodley: He denies the reification of that proposition.
Dan: Does that work? I’m skeptical. Turn propositions into objects (= reify). And then John tried to sing => John attempted the act of singing.
Woodley: A lot of like nominalization_rel, but everywhere.
Dan: Relative clauses: The dog that barks sings
Woodley/Francis: That’s just conjunction, not a problem.
Dan: Dogs that bark sing. How do you get the subset relation?
Woodley: For all (x) dog (x) and bark(e1, x) => sing(e2,x); or maybe it’s an existential. Might even be ambiguous. Relative clauses are not the problem. More like scopal adverbs etc. that get interesting.
Dan: Isn’t it the case that you have a limited set of primitive arguments? What do you do about The dog probably barks?
Woodley: Reify the barking. It’s not beautiful. The question is what do you want to do with this?
Francis: Or modal operators.
Woodley: If you don’t want to stay with FOL, can go to modal logic etc. Not necessarily first-order that’s the problem. It’s not that what we want to use can’t exist, it just involves elements of lots of logics developed separately.
Woodley: These are all logics that have research on them and that you can reason over, in a way that you can’t do for MRS.
Francis: In the Kyoto project I was in, people at the level above took MRSs annotated with WordNet senses and put into Dolce ontology (search application). Not sure which interpretresting things couldn’t be done, but some stuff could be done.
Emily: What’s the point? Bringing a linguistic, broad-coverage perspective to questions of natural language semantics.
Woodley: I don’t know that we’ve done this. We haven’t taken the ERG/Redwoods and said here are these 15 things that are hard to do in logic.
Emily: That’s what the ESD project is about. Documenting what’s there, so we can ask later what’s hard in any given object language.
oe: Motivates the distinction between description and object language, because we have the freedom to use the description language to work with to turn this all up, while the object language isn’t a solved problem.
Woodley: Like building a house: need to figure out what rooms, etc before laying the foundation.
oe: Can even leave open the question of whether that foundation has to be build from concrete (logic) or something else (something functional).
Emily: Or reinforced concrete (logic++). Like: information structure, connotation.
Dan: cynicism?
Emily: So have we answered the question?
Woodley: Provided some candidate answers, not all of which I was expecting.
Dan: We’re pretty sure that the people who do logic can help with the really large question we’re asking (how do people use language, anchor it to the real world), but we’re a pretty long way from what logic is currently doing or able to do. Part of our endeavor is to clarify that specification of work. More clearly articulate what the short-comings are of the available platform for logic.
Francis: One possible way towards that is to try producing some logical form and see what we lose.
Dan: Can probably make any of these platforms do rather more dancing around than you might think. (Though it might be ugly.)
Woodley: MRS describes the relationship between what the grammar produces and a particular object language. The object language of the ERG is not FOL+reifications.
Emily: Not mappable.
Woodley: Mappable. The Copestake et al 2005 describes a process for mapping to a particular object language. (Scope-resolved MRSs.)
Emily: You think the a scope-resolved MRS is a logical formula? That might be true for those in the MRS paper (oe: maybe not even that) but not at the schedule for the full ERG.
Dan: Underspecified MRS for Almost but not every dog barks doesn’t resolve to any known logical form, except if you have a predicate almost-but-not-every
Emily: Not interpretable.
Woodley: No, but still a formula.
Emily: So that’s not an interesting object language.
oe: Wait, I need some more terminology:
- UDMRS
- MRS
- scope-resolved MRS (SRMRS)
- object language
Woodley: SRMRS == OL?
Emily: Can you call it (an SRMRS) a logical formula if it doesn’t have an interpretation?
Woodley: My understanding is that the mapping to a fully resolved MRS was meant to be a mapping to something that corresponds to an OL formula, and if there isn’t an interpretation of that formula, then that’s a flaw in the ERG (naively).
Emily: I’d say it’s a flaw in logic.
Francis: (quotes from correspondence theory). If we don’t have a semantics for the object language (target language), is it an object language?
Dan: SRMRS is an essential step to mapping to any OL.
Woodley: They can have names like almost-but-not-every-q.
Dan: No, not if you’re going make me a OL robust enough to handle what the ERG can describe, because these are open ended..
Woodley: You’re making the assumption that FOL + GQ has a closed set of quantifiers.
oe: There’s a strong tradition of that.
Emily: Because they all need special treatment in the model theoretic interpretation.
Dan: If we came to agreement about the terminology, what would it lead us to?
Woodley: Quotes from MRS paper: MRS is not in itself a semantic theory but can be most simply thought of as metalanguage for describing structures in some object language.
Dan: for describing == can be used for.
Francis: MRS output gives us formally well-structured ?algebra/?formal thing, but until we can link that to something interpretable just saying is ‘all’ and ‘but’ and ‘not’ and ‘every’ and these are connected, it’s not the description of the world we’re aiming at. It is the case that it is a coherent consistent system of a description.
Dan: It provides some constraints on what we express.
Francis: I think we all agree that the straight output of a scope resolved MRS is not the final target of what we’re trying to do.
Woodley: I think it’s an open question whether the right approach is to translate that to something else that has an interpretation, or to assign an interpretation directly to that.
oe: Then I would no longer hesitate to call them formulae.
Woodley: I always thought that was the long term goal, but it sounds like what everyone else thinks it would be better to map to something to that has interpretations already.
Dan: Speaking for the authors, that seemed like the smarter way to go, to build on 2000 years of work.
Woodley: But that is still consistent with my interpretation of the paper.
Dan: We never intended to do that interpretation ourselves. That was never a design goal for the MRS, to produce something that would sustain direct interpretation. That did give us license to be lazy, and to push the interpretation off to someone else. We still do that, we glue things together using the syntax of the language and hope that it can map to something that is interpretable.
Emily: So to rephrase the earlier task: if we try to map to some object language and find some things unmappable, flaw might be in the logic or it might be in the ERG.
Dan: We may find that there’s a better way, e.g., to deal with variable arity.
Dan/oe: He already ate. if that means you have to say so in the predication.
Emily: Do we need more input from outside (task-based)?
Woodley: Do you have a mechanism for taking sentence from Etchemendy book and outputting one of their logical forms?
Dan: Yes.
Emily: For a very small subset.
Dan: Not perfect in the mapping from language to logic.
Emily: If what you want to say is things that logicians say about a block world, it’s possible to get from ERG to propositional logic.
Dan: Yes. And because there’s a reasoner, can see if they’re equivalent.
Francis: But I put the block nearly but not quite on the stack.
Emily/oe: What you can say about the block world is part of the block world.
oe: Reading from the paper: goal is semantic representations. My interpretation through the years has been that we don’t know what the object language will be, but that that shouldn’t keep us from working on the representations.
oe: ‘the object language’ is probably misleading here.
Emily: It’s definite/non-specific.
Dan: It doesn’t exist.
Woodley: There could be MRSs that describe sentences of FOL, there could be MRSs that describe sentences of modal logic, of SQL etc.
oe: And all of these could be viewed as within ‘predicate logic with generalized quantifiers’
oe: predicate logic with GQ doesn’t denote one language; it’s a family of candidate languages.
Francis: I think it’s more important to talk about what we want to do, rather than on what the authors of the paper meant.
oe: Harry Bunt started to work once on a model theoretic interpretation of MRS.
Emily: What happened?
Dan: He retired.
Woodley: Summarizes ‘the point’:
- Facilitating applications like MT, search
- Maybe more surprisingly inform the logicians.
Emily: Not surprising: linguistic discovery has been part of the point all along.
Woodley: Syntactic discoveries maybe more expected that input to logicians.
Woodley: With that in mind, maybe can go back to some of these questions.
oe: Paper “Semantic underspecification and modifier attachment ambiguities” Egg and Lebeth. Looks like they’re attempting something similar to Dan/Woodley discussions. Also a section on the interpretation of MRS.
Woodley: Ooh! lambdas!
[ Reading that paper; Emily leaves room momentarily ]
Woodley: I approve of clean specs; I’m less concerned about the other person.
Francis: I think it we do the second, we’re more likely to come up with something to people, then if we just try to come up with something that’s appealing to people. We’ve produced the most beautiful trees known to man, but no one uses them because we haven’t shown them how.
oe: I think the inform the logicians goal could be more foregrounded, but we’re a little short on tame logicians.
Dan: Tell me more?
oe: I think ESD could turn into something there. If we put together out test suite with what we think are defensible representation, we could approach peers and some of them could be logicians and solicit feedback — that second workshop in the series.
Dan: Not really talking to logicians.
oe: That depends on who we invite.
oe: To me, the what’s the point question has been satisfactorily answered.
Woodley: There are some good answers there, but it’s not necessarily complete.
Last update: 2014-02-16 by EmilyBender [edit]