Skip to content

SofiaIconsImplementation

StephanOepen edited this page Jul 5, 2012 · 3 revisions

Emily & Sanghoun's initial proposal:

In general, the idea is that for any given pair of indices, if there is an ICONS element in the MRS associated with a realization that realization is valid iff:

--- There is no ICONS element associated with the same pair of indices in the input; or --- Any ICONS element associated with the same pair of indices in the input is compatible in its type with the ICONS element in the candidate realization.

(I'm not sure as I type this what to do about cases where there are multiple ICONS elements for the same pair of indices.  As far as information structure is concerned, we wouldn't expect to have more than one.  Of course, if we're putting other stuff into ICONS, then there might be ICONS constraints of different types. All of the above would then be only applicable to subtypes of info-str.)

Refinement: all the variables in the ICONS should be variables (?characteristic variables: Emily says no because of pro-drop, esp topic drop in Chinese) somewhere else in the MRS? Is it ever desirable to say "this element is a topic, but we don't know of which clause?"/"This clause has a topic, but I don't know what it is"? Jacy has something like this just now, just because in that particular construction can't grab the relevant clause variable. Start with all variables are somewhere else in the MRS, and then look into that further later.

Bracket the question of whether the [e2 info-str x1] even belongs in the ICONS list from how to handle the ICONS list in generation. More general is to allow for both possibilities, and not assume that one or the other will be the final solution.

Absence of specification meaning something (lack of equation of variables in MRS) vs. underspecification. These two things cannot be equated in realization. Procedurality? Notion of default?

Do we need a notion of defaults in dealing with ICONS? Notion of defaults: avoid generating topicalized sentences by taking underspecified as e.g., non-topic. Could be done via input MRS fix-up transfer grammar (provided support for ICONS).

Could similarly do default inequalities for the variable thing, but wouldn't want to implement that. That's the way to get around the wince/whinge-inducing "procedural" term.

EB's proposal regarding enhancement of current post-generation compatibility test with notion of ICONS subsumption. Do this generate and test as the first implementation.

FCB's request for variable-property sharing works differently. Where does the translation between that special icons subtype and actual variable property sharing happen in the generation pipeline?

Might be tempting to look at ICONS as parallel to HCONS, but currently trying to do better than that. HCONS is part of the idiosyncrasy of our notion of subsumption between MRSs. Cheap scope - equates qeqs to produce a normal form ordering of the quantifiers all at the top, equating skolem constants but not the variables. Current standard post-generation test doesn't see HCONS. Might be dependent on only using leqs. Third variant tests exact isomorphism including HCONS; less efficient.

qeqs easier than leqs: every time there's a qeq it corresponds to a scopal argument, makes the transitive closure trivial. In the course of refining the grammar got rid of cases where we were doing different things with the scopal arguments (e.g., equality). Assuming actual identity works for a particular grammar, but not necessarily for generator input (makes it harder to construct input to the generator). German needs leqs, probably also Turkish...

Upshot: not many useful parallels between HCONS and ICONS

Do ICONS have logical status? Not for information structure (probably), maybe for anaphora.

So coreference:

Two types (maybe): coreference with variable property sharing, coreference without constraints on variable property sharing. There might be more fine-grained things we want to do about how much variable property matching is done.

Lars: Introducing semantic roles into the grammar => need to detach number from coreference.

Do we ever get "coindexing" (variable property sharing) without coreference:

Elephants are mammals.

Do want the number to match (some of the time), but don't want to say they are coreferent.

These five people are a team.

-> Suggests that this not a grammar property, but rather something that we should expect in the input MRS.

But the input MRS is often underspecified for number in e.g., jpn > eng translation.

What about:

She grew up a healthy child/*healthy children. (Plus also gender agreement on child in Russian.) But don't want to make these coreferent in a real-world sense? How would we test that empirically? Since we're talking about different life stages, saying "same entity" may or may not make sense.

What about negation?

She didn't grow up a healthy child. (Still need agreement in Russian) Elephants are not mammals. (But too much mismatch in the identity copula.)

Maybe this is grammar internal; doesn't necessarily need ICONS.

There's two children and they grew up a happy pair. They went through life a happy couple.

-> Number's not agreeing here.

What can be done about passive in English? There is no direct relationship between passive and information structure.

What was destroyed by the earthquake? Three buildings were destroyed by the earthquake.

Who was this book written by? This book was written by Kim.

EB is interested to know of the examples where the need is felt to control active v. passive.

Maybe we could have another ICONS type that talks about promotion of demotion of arguments. (Which potentially interact with information structure.)

Summary of TODOs so far: 
-- MRS object augmented with representation of ICONS (Ann)
-- View these (Ann, oe)
-- Serialize (xml, simple output and input)  [output in PET later after we've finished experimentation in LKB]
-- Post-generation comparison (Ann adding function on top of existing code, oe will improve later)
-- Add subsumption hierarchy to the SEM-I (Sanghoun); but Ann might hack^H^H work around this for now and let someone else fix later.

What about non-restrictive relatives or appositives?

Rather than appos_rel (which is problematic because its handle doesn't have a place to live in the scope tree), use ICONS:

My brother, the doctor, sings.

Where does the doctor go in the scope tree, if appos_rel isn't connecting them? doctor is down inside the restriction of the, the being a quantifier can attach high in the tree and pulls doctor along with it.

My brother who sings dances.

Non-restrictive reading should have same or similar to semantics; restrictive reading has it inside restrictor of my, non-restrictive doesn't.

Building the compositional semantics where we always do, but rather than having the apposition rule add new EPs, add an ICONS showing coreference or equality. Is that the same name as the one we use for anaphora? Name it appos only if it's different semantically from what's going on. But what are the examples that we could check against?

Note that predicative noun phrases are a hard, unsolved problem. (Ann observes that this is a case where composition and formal semantics point to different answers.... maybe ICONS could have some part in the solution.)

How should the generator use this information in the ICONS? If the grammar is suitably updated, should see apposition iff the input MRS has the ICONS. Oe points out significant potential cost in efficiency of generation, since currently the appos rule is only triggered when appos_rel is in the input semantics because of C-CONT.

Ann: put this under the heading of stage 2. Engineers (Woodley) think there may be an efficient way to do this.

Dan fantasizes about the padded cell where grammar engineers don't have to worry about efficiency. oe briefly fantasizes about moving all variable equality into ICONS. Ann tops that by suggesting she'd like to move semantic composition out of the feature structures.

Objections:

Kim tried to sing. --- there's no room for slipperiness in the identity between the trier and the singer?

Sounds cluttered. What's the advantage?

Might eliminate the formal idiosyncrasy we currently have with the treatment of variables.

Could apposition be done like coordination --- creating a new index? -> More quantifiers in the semantics, since we need one more for that index.

For the number agreement tendency cases when the input doesn't have the number but we can't make a hard constraint, could we find a way to define features for the realization ranking machinery that can capture this information. (Also possibly related to notion of probabilistic connection between passive and information structure.)

Do ICONS affect truth conditions?

  • Partee examples showing that information structure does sometimes interact with truth conditions.
  • Other examples showing interaction with prosody and truth conditions.

If ICONS doesn't make it onto a scope-resolved LF, how can we model its impact on truth conditions?

What's the difference between ICONS and RELS (and HCONS), so we can tell when to use which tool?

  • ICONS are always binary relations between two individuals.
  • ICONS don't have/don't merit a position in the scope tree --- in fact mess things up when you try to put them into the scope tree.

Can we sharpen up the second of those? Can't be fit in naturally? Don't have a reason to be there? What about the n-n-compound (two-place unspec_rel). n-n-compound only sometimes has to do with identity.

oe: In what way are unspec_rel required to be in the scope tree? Ann: Easier to see in the preposition case, but also true in the compound case

Most white cats are deaf. Most deaf cats are white.

Need to know where white goes wrt to restriction to get the right truth conditions.

Most cats from Africa are in Bulgaria.

... resolution: it's not that we don't need to pin it down in composition by attaching handle somewhere but that we don't want them in the tree in the end, either because doing so would break the tree for other things, or because we don't have a natural position in the tree and have to pick arbitrarily.

My brother who sings dances. (non-restrictive interpretation)

If we're going to use ICONS to set up the relationship between the relative clause and the head noun (like appositives). Where does the relative clause attach in the scope tree if we do that? Dan hopes: there's enough quantifiers to make that work. Might need a non-restrictive who to introduce a new quantifier. Ann will look into this and report back.

Clone this wiki locally