TheAbbey_Chrysalis2014SchrodingerMrs

Schrödinger's MRS

Francis: Looking for one packed representation of the Secretary of State the compositional phrase and the entity.

oe: Not ambiguity, but vagueness?

Francis: Neither --- both at once. Quantum.

Woodley: Can't it just be both?

Francis: Knowing that they overlap could help with the generation task, so we don't generate Secretary of State twice. Want to mark one as main one as sub/one as opaque and one as transparent. Maybe CFROM/CTO will tell us they overlap.

Woodley: Probably always want to get rid of the opaque one for generating.

Francis: Do we want to mark things as transparent or as opaque, or just the opaque.

Woodley: I think everything is presumed transparent unless marked otherwise. Interesting question is how to relate the opaque ones to the bits of the transparent ones…

[ Topic picked up again later in the morning ]

Francis: Back to the modest proposals and overlapping representations co-present in the semantics. Do we have a good way of distinguishing them, or can we just wait for you to come up with one.

Ann: In what sense of distinguishing them?

Francis: Secretary of Defense: done compositionally, and the gazetteer says there's an analysis of that as a single entity.

Ann: Yes, parallel parts of the semantic lattice or something. If you think about this in general, you can always take all of the different analyses that the different analyses from the forest, and pack them into a single lattice.

Francis: In that case, my interpretation is not that two of them hold at the same time, but that I just haven't made a decision as to which one I'm selecting.

Ann: Yes. I'm suggesting giving up on single semantic output for single syntactic analysis assumption. But how to distinguish between them? In some cases, the syntax will distinguish between them.

Francis: Which them?

Dan: Those two candidate analysis.

Francis: I think we're arguing that they're both true simultaneously.

Ann: I agree that they're there simultaneously, and in many cases you don't need to choose.

Francis: If I'm generating, I'd like to generate only a single S of S.

Ann: If we're getting these things in the case where they're producing the same strings, can arbitrarily choose one or the other of the subgraphs to generate from, and it won't make a difference. Could work mechanically.

Dan: Do we get a differentiation if we look at the conjoined one: The Secretaries of State and War. In that case I can't use a single entity,

Francis: … unless we do some extra processing.

Ann: Do we get the entities out of there?

Dan/Francis: We were hoping to do so, via some additional magic. Going to need two secretaries.

Francis: If we're doing NEs through trigger rules, then we have mechanisms for forking.

oe: By moving NER to the MRS level, we thought we'd have a better handle on doing this kind of thing.

Ann: Will have to think about that example. Not sure I can see how the parallel analysis idea will work in that case.

Dan: In practice, we'll need some as-yet specified additional machinery in order to be able to hold that dual view. In the near term we'll have to fall back to the decomposed representation for generation.

Ann: Might not get the most compact one.

Francis: It's not wrong to generate that.

Woodley: It's more helpful than the user asked for, getting to The Secretaries of State and Defense if they gave you S. of State and S of War.

Francis: I don't want to generate The S. of State the S. of State left.

Woodley: You won't, it's ungrammatical.

oe: There's only one x.

Francis: And it's not a problem to have two EPs with the same char variable?

oe: Currently it is. We're talking about a bright and rosy future.

Francis: Checking that the bright and rosy future is well-formed.

Dan/oe: What about the S of State of State.

Ann: Two parallel subgraphs/lattice-y thing. Pick one and generate.

Woodley: Not yet clear where the opaque version comes from in parsing.

Francis: My view is that the current idiom machinery could be adapted to throw in a new predicate (rather than just raising a flag).

Woodley: In that world, the input could be either way, but by the time it goes into the generator, you'd have to get rid of one.

Francis/Emily: Usually the opaque version, unless that's the only one.

Ann: Good to think of this in conceptual terms before worrying about where the gazetteer comes in. You do have a lexical entry for S. of State there, and you might want to avoid doing that in the ERG by putting it in the gazetteer, but it's still a lexical entry in the broader sense.

Francis: I think there's still a difference from what we're currently doing in that we're saying not choosing between two, and keeping both.

Ann: Yes, that's new, and that's why we want to think about it in those terms, before looking at how to implement.

A Modest Proposal

By Ann (not an Elk). See also: A modest Proposal for Proper Nouns

I believe there are some cases in the list where underspecification is not feasible. I can explain that in more detail if you like. The alternative is to think of something where multiple analyses may be packed into a single structure. Under some circumstances, both interpretations may genuinely be available, in others, there will be a potential for disambiguation.

I've been thinking about this for some time - the case I've thought about most is semi-compositional MWEs. If we take an example like `red deer'

a) an ideal lexicon should record it as a separate lexical item (cf `elk')

b) when used of the particular species, it's still somewhat compositional - it's a species of deer, and is reddish-brown.

c) `a red deer' could be used fully compositionally - e.g., to refer to a scarlet toy

d) `a very red deer' could be said, but wouldn't be analysed in terms of the MWE semantics (one would have to say `a very red red deer')

e) `an albino red deer' can only involve the MWE - it isn't a contradiction, but does wipe out any implication of redness

I don't see much wrong with having a dual structure - e.g., in DMRS terms, a representation containing both:

red_a_rel --> deer_n_rel

• ARG1/=

red_deer_n_rel

One could perhaps mechanically achieve this in a grammar by having a dual semantics in a single lexical entry for `red deer'. Aurelie mentions `black key' on a piano as another case - they aren't necessarily black, but someone who didn't know the MWE would nearly always be able to treat this as compositional. However, it's an idiom of encoding - they could be called sharps and flats.

Compositional but established compound nouns show a slightly different issue. If you believe that garbage man' has a use in context man made out of garbage' (analogous to `snowman', as argued by someone - can't remember who at the moment) then we need both the established MWE and the fully compositional general compound_rel reading. Here, underspecifying doesn't make sense - the general compound_rel use is fully underspecified, but if we just use that, we have no way of recording the established use. I guess this doesn't matter so much for the current ERG, but it's a problem if you believe the system should be capable of representing facts about linguistic conventionality.

Anyway, I don't see why one shouldn't extend this idea further and use it e.g., in cases like `baked a cake for Kim'. If you want to represent the similarity with `baked Kim a cake', then why not have both the ARG3 and the preposition there?

Having a distinction between:

`gave Kim a cake'

and

`gave a cake to Kim'

in that the latter has the preposition as well as the ARG3, might be useful for the

`gave Kim a headache'

*`gave a headache to Kim' puzzle than Georgia Green wrote about