I currently cannot attach a focus particle to a wh-word because my wh-word is [ LIGHT + ] and the particle from the infostructure library says its MOD is [ LIGHT - ]. I see in Sanghoun’s thesis that [LIGHT + ] indicates a “word” while [ LIGHT - ] a phrase. I am not sure what that means. Is a bare NP like who a word or a phrase? For some reason which I am still investigating, my who NP is [ LIGHT + ] but Ivan NP is [ LIGHT - ].
Does your wh pronoun go through a bare-np rule? That might be what’s making the difference…
Both Ivan and who are bare NPs in my two test sentences (who FOC arrived? Ivan FOC arrived). There is the difference that Ivan first goes through a lexical rule while who.NOM is fullform in the lexicon.
Oh, I think I see what you mean. who appears as a NP but it didn’t actually go through it! How do I make it to?.. Or should I hardcode [LIGHT - ] for it in the lexicon?
Something in here apparently means the word is of type lex-xynsem which is [ LIGHT + ]. I can’t trace it yet though:
wh-word-lex := norm-hook-lex-item & basic-icons-lex-item & [ SYNSEM [ LOCAL [ CAT.VAL [ SPR < >, SUBJ < >, COMPS < >, SPEC < > ], CONT [ RELS.LIST < [ LBL #larg, ARG0 #arg0 ], [ PRED "which_q_rel", ARG0 #arg0, RSTR #harg ] >, HCONS.LIST < [ HARG #harg, LARG #larg ] > ] ], NON-LOCAL.QUE.LIST < #arg0 >, L-QUE + ] ]. wh-pronoun-noun-lex := wh-word-lex & non-mod-lex-item & zero-arg-que & [ SYNSEM [ LOCAL [ CAT.HEAD noun, CONT [ RELS.LIST < [ ARG0 ref-ind ], [ ] >, HOOK.INDEX.PNG.PER 3rd ] ], LIGHT - ] ].
You don’t want the wh pronouns to go through bare-np because they already have their quantifiers (which_q_rel); giving them a non-empty SPR value would mean they have to go through bare-np, and then they’d get a second quantifier (exist_q_rel), which would mean broken MRSes.
You’re right that lex-synsem is likely the culprit. As for figuring out where it’s coming from, I suggest using the LKB to look at the expanded type for each of those supertypes. One or more will (probably) have [ SYNSEM lex-synsem ] and then you can trace backwards from those to see whose introducing it.
OK, here’s what I did:
- Looked at expanded type in the LKB starting from wh-pronoun-noun-lex and then each of the supertypes in turn, for as long as I was seeing lex-synsem in the type: wh-word-lex, norm-hook-lex-item, norm-ltop-lex-item.
- norm-ltop-lex-item is still a lex-synsem while its supertype is lex-item and that one is a synsem.
Does this mean norm-ltop-lex-item is the “culprit”?
norm-ltop-lex-item := lex-item & [ SYNSEM [ LOCAL.CONT [ HOOK [ LTOP #ltop ], RELS.LIST.FIRST #keyrel ], LKEYS.KEYREL #keyrel & [ LBL #ltop ] ] ].
How do I deal with it now? If I simply copy the same constraints, I imagine the effect will be still that my wh-pronoun will remain a lex-synsem. But then the constraints are meaningful (in the area which I still don’t fully understand).
I re-created the type for wh-word without using norm-hook and norm-ltop lex-item types, copying all the same constraints manually:
wh-word-lex := basic-icons-lex-item & [ SYNSEM [ LOCAL [ CAT.VAL [ SPR < >, SUBJ < >, COMPS < >, SPEC < > ], CONT [ HOOK [ INDEX #arg0, LTOP #larg ], RELS.LIST < #keyrel & [ LBL #larg, ARG0 #arg0 ], [ PRED "which_q_rel", ARG0 #arg0, RSTR #harg ] >, HCONS.LIST < [ HARG #harg, LARG #larg ] > ] ], NON-LOCAL.QUE.LIST < #arg0 >, L-QUE +, LKEYS.KEYREL #keyrel ] ].
This of course is still a lex-synsem, as it is the same structure. But without the LTOP-KEYREL identities, I do not get the pronoun relations.
After further conversation with @ebender, it looks like what should be changed here is not the wh-word but instead the particle. A particle of this sort attaches to words, not phrases, so it should require MOD < [ LIGHT + ] >.
So if I have a focus particle which I want to attach also to question words, I seem to need to say that the particle’s MOD element is LIGHT +. However that leads to ambiguity with regular common nouns modified by the same particle: now the particle can attach both below or above the bare NP rule. That, I think, was the point of having them require [LIGHT -] on their MOD in the first place…
If I say that in such languages the bare NP rule is [LIGHT -], this goes away. But is that meaningful?
Yes, I think it’s fine to say that the bare-np rule is [ LIGHT - ], and in fact I’m rather surprised that that isn’t already the case.
Do you mean, in matrix.tdl? If I add that to basic-bare-np-phrase, some tests (including old tests, so, not just my new tests) fail (I could investigate why, one day, but perhaps not now; it could be that actually some unwanted ambiguity goes away).
Hm, that is surprising. Maybe it would be good to file a ticket on this one, and then look into it as a bug this summer But it should be fine for your specific case.
So, this is related to this (to how to fix or not fix this, maybe): in languages where there are determiners, we expect the modifiers to be able to attach before the determiner, is that right? The beautiful cat and so on. So, with PPs, it should also unambiguously be [The [cat in the house]], right? Ideally there won’t be [[The cat] [in the house]]? (For an NP; I am not talking here about the I saw the cat in the house type of ambiguity).
That sounds reasonable for English, but I don’t think that’s cross-linguistically true. Modifiers attaching after determiners was one motivating example for a talk I gave in Oslo.
Thank you, Guy! I am trying to understand the slides (too bad we don’t have a recording!) The IGT there is about verb modifiers (and it is clear there how negation must attach below the adverb), and then the final abstract example has a noun and a determiner, but I can’t quite decipher what the slide is saying.
Yes, I could have prepared notes to accompany the slides…
Most of the slides are about non-scopal modifiers attaching after scopal modifiers. The last slide is saying that we have a very similar problem and solution for non-scopal modifiers attaching after quantifiers. In both cases, in the MRS, the non-scopal modifier MRS needs to share a label and take an argument. In DMRS, this is simple, because we just have an ARGn/EQ link to the INDEX node. In MRS it’s awkward because the corresponding label may no longer be available in the HOOK.
I didn’t give examples in the slides… I think there are examples in Turkish? Perhaps @ebender can remember!
We talk about some examples here: http://faculty.washington.edu/ebender/papers/ben_egg_tep.pdf
BUT: The key thing to note is that with the existing types in the Matrix, allowing modifier attachment to NP is going to give broken semantics.
Actually, conceptually, why would bare NP be [ LIGHT - ]? Or in other words, why couldn’t we expect it to be $+$ (e.g. in languages where nouns don’t get articles)?
[ LIGHT + ] means ‘can show up where only single words can show up’ … I guess I would be surprised to find a language that has a special position for single-word NPs (rather than say a special position for pronouns only).