Skip to content

SaarlandLtopDiscussion

EmilyBender edited this page Aug 4, 2013 · 7 revisions

Discussion on constraining LTOP on NPs in MRSs

Leader: Emily

Participants: Joshua, Stephan, Rebecca, Lars, Dan

Potential examples where we might want LTOP available after composing an NP:

  • English appositives, if we want the appos_rel's LBL to be identified with the head rel of the left NP.
  • Turkish N-Det-Adj
  • Yiddish Intens-Det-Adj-N
  • Latin Adj-P-N (e.g. magna cum laude, imo a pectore, medias in res)

Obstacles to making the LTOP available: Semantic algebra prohibits binding LTOP of a quantified NP, since the LTOP of an argument should either be shared with that of the head, or be a scopal argument of the head. This sem-alg constraint is useful if we dispense with quantifiers for pronouns, since the pronoun's non-q relation's LTOP needs to be identified with that of its functor. Similarly for our treatment of apposition in the 1212 ERG.

Options:

  • (A) Add a unary rule that empties the LTOP of a quantified NP (which starts out with the LTOP bound to the head noun's LBL, to sustain further intersective modification), and liberates it to be an argument. Currently implemented in the ERG to enable e.g. |everybody in Paris who dances|, which would be problematic since |everybody|.
  • (B) Give up the semantic algebra constraint (abandon strong hypothesis), and allow LTOP of NP to be identified with the head noun's LBL. Potentially problematic if we give up quantifiers for pronouns, as has been contemplated for awhile for the ERG. We could go further and make pronouns introduce no rels, and then coerce them to add a nominal rel for |you in the corner|. But without a quantifier, we would have the LBL of the preposition identified with that of the head verb in |you in the corner sing|. Not appealing, since it would be a first instance of an intersective modifier whose ARG1's source relation does not share its label.
  • (C) Add another attribute in HOOK for NPs to keep track of the head noun's LBL. Weakest, and potentially problematic downstream.

Notes from the bonus scribe:

Dan: Two cases standing are Turkish and Yiddish (zeyer), maybe introduce the quantifiers in non-branching NP rule even if we see something that looks like a determiner. (cf. Emily's earlier points about disassociating demonstratives from quantifiers in languages with adjective-type demonstratives.)

Dan: LTOP of the argument is interesting in two cases:

Directional PP arguments Pronouns introducing just a pron_rel

  • -> Need to obey the algebra so that the pron_rel is bound to something (namely the selecting head)

Not yet done in the ERG, but ready for it (probably)

Emily: Could take the further step of removing pron_rel as well as pronoun_q_rel.

Dan: Not necessarily opposed to that; would allow revising our view of the algebra.

Dan: Previously appos_rel was also a case where this was required to make sure it gets bound to something. Stripes the cat meowed appos_rel was identified with the empty LTOP of the NP, but then the algebra unifies that with the head's LTOP (per algebra), hooking it in. That problem should go away as we move to ICONS, but still another place where the semantic algebra constraint is doing some work.

oe: Not sure ICONS fully solves the appos problem (see interesting unanswered questions about ICONS and truth conditions yesterday), so it might be premature to move the appos problem off the table in this context already. If we had the label of the other N' available, that would also solve the appos problem.

oe: Another potentially heretic idea is to put the N-bar's LTOP somewhere else in the HOOK.

Dan: Tried putting it in the XARG previously, given that NPs don't really need an XARG usually … but that doesn't really turn out to be true. There are much more legit uses of XARG in NPs, e.g., gerundives my singing of the song --- keep the XARG from the verbal relation.

oe: Would have to introduce a new attribute (e.g., ITOP - internal top) or deviate from the algebra.

Lars: Norwegian has examples like:

  • the nice boy-the
  • boy-the
  • *nice boy-the

… one of the determiner-looking things has to yield to the other when both are present.

Lars/Dan: everybody in Paris is also a problem in the English the grammar. Want this to mean every(x), person(h,x), in(h,e,x,y), Paris(y)

oe/Emily: So what do you do?

Dan: Have tried several things. At present special treatment of that small class of lexical items as entities that select for semantic modifiers as (optional) complements.

oe: Not currently: looks like the pumping rule

Dan: Right --- attach the quantifier low but keep the LTOP exposed in a constituent that isn't ready to play in the syntax except for participating in intersective modification. Non-branching rule removes the LTOP (rather than introducing the quantifier).

Dan: The problem with the complement approach is that it doesn't scale to multiple complements: Everybody in Paris who bought a cat.

Emily: Clever --- is this algebra-compliant? But first: how do you keep it from playing in the broader syntax?

Dan: Probably by exploiting some version of the anti-synsem mechanism of the SPR list.

oe: To paraphrase it has the distribution of the N'?

Dan: No -- can't conjoin with N's, can't appear in N' appositive constructions. It's a quantified NP that can still undergo further (intersective) modification.

oe: And unary rules are not subject to the algebra. What that unary rule does in terms of the semantics … but the algebra doesn't say anything about the semantics.

Emily/oe: Unary rules that are semantically empty --- so unary from the point of view of the semantics as well.

oe: To judge this as algebra-compliant, need to say that the algebra has nothing to say about a certain class of unary rules.

Dan: flies under the radar.

Emily: might work for Turkish. But predicting no non-intersective modifiers of nouns?

Dan: More subtle than that: Predicting no non-intersective modifiers that need to be picked up in these cases. So not a very strong prediction.

Emily: I think English is maybe more cooperative in this case than something like Turkish might be. But currently no non-intersective modifiers in ERG?

Dan: No --- still waiting for compelling evidence. Just the intensionality doesn't seem enough.

Emily: What about likely, probable etc as in the most likely winner of every medal was disqualified?

Dan: Jan Tøre and others have pointed out that there is more than two classes of what happens with adjectives. Not just scopal + intersective. … probably better to think in terms of languages that make the problem clearer

[ Emily gone ]

Hypothetical option B making the LTOP on NPs the N-bar's LTOP. What challenges does that lead to?

oe: Does that force Dan into two HCRs, but it doesn't:

Dan: Split both HSR and HMR for scopal and non-scopal dependents. Haven't done that for HCR, that's done in the lexical types instead.

oe: Biggest challenge would be having to declare oneself openly in opposition to the algebra.

Dan: Not so troubled by that, but troubled by the tempting strategy of throwing away the quantifier of pronouns. That requires the algebra.

Emily: Unless you also toss the pron_rel.

oe: That would be problematic for the variable-free views on the semantics (EDS or DMRS), no EP introducing it as its inherent variable.

Dan: Also problematic for sentences like You in the corner should buy a book: label of in?

Emily: Wambaya: Bender 2008 concludes that Wambaya facts are compatible with DTFS but not the algebra. Do have modifiers of otherwise unexpressed argument positions … but not sure what I'm doing with the label of those quantifiers. (Maybe introducing quantifiers in the grammar currently to tuck the modifiers under, but no pron_rels --- so the opposite solution.)

oe: To further illuminate the LTOP on NP problem, need a definitive conclusion on the pronoun space.

Lars: Why remove the pronoun_q_rel?

Dan: Can't find a semanticist who wants it, except possibly Alex.

oe: Remember in Cambridge II being convinced that it was on its way out. But after Hankø thinking it was sticking around. What are the arguments in either direction?

Dan: For getting rid of it: Occam's Razor only.

Lars: Why posit in the first place?

Dan: Because we went with the rule that all nominal indices need to be quantified. But then semanticists said that's weird, because pronouns aren't like regular noun phrases that need to be quantified.

oe: Effectively introducing unquantified referential indices.

Dan: Unquantified indices of some kind.

Lars: What about proper names?

Dan: Semanticists are less unified on this one, because they clearly have both uses. Pollard says that a proper name like Kim is a constant, but can be coerced into common noun uses. We've even talk about clever strategies for lexical entries for proper names underspecified and then have the syntax decide.

oe: Do we need a new type of instance variable?

Dan: Not for the external MRS, but maybe internally in the TFSs.

oe: Some of the scope rules are built on the assumption that all ref indices need to be quantified. They'd need to be something like unquantified constant indices...

Lars: … clitic doubling …

oe: admits that maybe the variable-free semantics isn't to be treated as hard constraint on design of semantic representations.

Dan (summarizing):

  • Option A: unary rule that allows the LTOP to be available for further modification, but protects the algebra by liberating the LTOP and syntactically moves from modifier target to argument-able.

  • Option B: give up the strong hypothesis on the semantic algebra, allow that LTOP to be the N-bar's label all the time. Just make sure not to grab it by accident. --> problems with anticipated move to dropping the quantifiers for pronouns.

  • Option C: adding another attribute in the HOOK (ITOP) … NTOP (noun's top handle) not really a TOP anything.

Dan: Don't care what the name is because I don't think we want to go that way. HOOK attributes have an engineering cost.

Emily: And a theoretical cost.

Dan: It doesn't help anything to say that we want to do two things at one time: want to have it visible (for modifiers) and not visible (for embedding functors). There would be some utility in thinking harder about this distinction rather than just throwing a feature at it.

Emily: Keeping the discipline of HOOK makes us think harder about these questions. In your grammar you've created a description of these cases via the rule and we could then do some bottom-up exploration of where the issues come up.

Dan: Yes, there are other cases: you who eat fish can still build while there's a quantifier for pronouns, with the same unary rule.

Emily: Does the pumping rule apply to all NPs or just these ones?

Dan: Just the funny ones.

Emily: Two objections Dan's (linguistic, about you in the corner), oe's (formal, variable-free rules).

  • you in the corner
  • *they in the corner
  • *he in the corner
  • those in the corner
  • we in this room like fish

[oe: okay in Norwegian even with third person pronouns] [Afterwards we noted that them in the corners sounds better than they in the corner.]

Emily: So are we dropping pronoun_q_rel on demonstrative pronouns?

Dan: Currently demonstrative_q_rel, but you've said that's not right.

  • you should stop laughing

  • you in the corner should stop laughing

pron_rel, in_rel

Emily: okay with you sharing label of should?

Dan: yes

Emily: in the second case too?

Dan: that follows

Emily: and in sharing its label with should?

Dan: not so much

Emily: I think the problem is the same regardless of whether we have pron_rel (as soon as we get rid of pronoun_q_rel)

  • you in the corner sing in a large hall
Clone this wiki locally