What to do with QUE (and YNQ) in wh-question filler-gap phrase

In the Grammar Matrix, we currently have several append-list-valued features: REL, SLASH, QUE, and, most recently, YNQ (for second position clitics in polar questions).

I am trying to finalize some kind of (unified?) account of all these features in the filler-gap phrase (or perhaps just in the wh-question subtype of the filler-gap phrase).

Here’s the basic filler phrase in its current form. As you can see, the append lists are mostly underspecified:

basic-filler-phrase := binary-phrase & phrasal &
                             SPR < > ] ],
    ARGS < [ SYNSEM [ LOCAL local &
			    [ CAT.VAL [ SUBJ olist,
					COMPS olist,
					SPR olist ],
			      CTXT.ACTIVATED + ],
		      NON-LOCAL [ SLASH append-list, QUE append-list, REL 0-alist, YNQ append-list ] ] ],
	   [ SYNSEM [ LOCAL.CAT [ VAL.COMPS olist ],
		      NON-LOCAL [ SLASH append-list,
				  REL 0-alist,
				  QUE append-list, YNQ append-list ] ] ] > ].

The wh-question phrase operates on the SLASH list as follows:

1st-head-filler-phrase := basic-filler-phrase & head-compositional &
    ARGS < [ SYNSEM.LOCAL #local &
                          [ CAT.HEAD +nrpd ] ],
           [ SYNSEM.NON-LOCAL.SLASH.LIST < #local . #slash >   ] > ].

The question remains what to do with the rest of the lists, particularly QUE and YNQ, which, when non-empty, normally trigger the application of a unary rule. In the ERG, I think they would be zeroed-out on the mother, and this way if a filler-gap rule applied, an in-situ rule won’t apply, etc:

; DPF 21-jul-04 - Removed parent head_nexus_que_phrase since this didn't allow
; multiple-wh clauses like "who hired whom", and it wrongly implied that there
; would be filler-head constructions where the mother still had a non-empty
; QUE value.

head_filler_phrase := basic_head_filler_phrase & head_final & binary_phrase &
             NONLOC.QUE 0-dlist ],
    ARGS < [ SYNSEM.LOCAL.CAT.VAL.SPR *olist* ], sign > ].

For a while, this is what I’ve been doing as well for the Matrix wh-questions library and for my Russian development grammar in particular. It mostly seems to work, though I keep thinking: if we take out things from SLASH one by one, why aren’t we doing the same thing for QUE?

So, while zeroing QUE and YNQ mostly works out, I have a slight problem with sentences like (1):

(1) Кто     знает,   где   мы  что      видели?
    Kto     znaet,   gde   my  chto     videli?
    who.NOM know.3SG where 1PL what.ACC see.1PL
Who knows what we saw where? [rus]

For this one, because the top node is licensed by a filler-gap rule (and the QUE list is zeroed out), I still get an additional tree with the in-situ phrase licensing the embedded we saw what:

Screen Shot 2020-04-29 at 10.21.47 AM

Any thoughts on this? Again, this is an additional tree to the one where where/gde is extracted and gets attached by an filler-gap rule.

I am already using MODIFIED on the mother of the in-situ clause to block extracting adjuncts from it. The mother is also L-QUE -, so, disallows a wh-word in its left periphery. But none of this helps when the question is embedded: the wh-word is no longer in the left periphery, and a regular adjunct-head rule may apply to attach the wh-adverb, so the MODIFIED constraint no longer helps, either. Would it make any sense to constrain the normal head-adjunct rules for MODIFIED? Probably not, right, because of course there may be multiple adjuncts.

I cannot address this by constraining normal adj-head phrase to not take wh-words as adjuncts when the adjunct ends up left-peripheral because then I lose some coverage (such as Who knows we bought what where, in Russian, or even just Ivan HOW arrived?).

All in all, seems like there is a lot embedded in embedded questions…

To answer my own question there, one reason this wouldn’t work currently is because our in-situ wh-words are not QUE-empty. So, unless every phrase structure rule “checks off” a QUE-value somehow… Doesn’t make sense.

I don’t understand what the question is here.

Isn’t REL for relative clauses, rather than questions?

What do you mean by “zeroing out” a list? These lists need to be dealt with one element at a time.

Spurious ambiguity due to free word order seems like a separate issue and not one to worry about here.

The question is what to do about ambiguity in (1) that arises because of the in-situ-question phrase + regular adjunct-head on the one hand and filler-gap + extracted adjunct rule on the other.

Maybe… Why are you saying it is due to word order? I wouldn’t have this problem if I didn’t set the filler-gap mother’s QUE to 0-alist (I would have other problems :slight_smile: ).

I mean saying it is 0-alist, without it being either identified with the head daughter’s or appended from both daughters.

You’d think (and I agree intuitively), but (a) look at the head-filler-phrase from the ERG above and (b) that would require a different analysis of in situ question phrases than we have now. More generally, this is my question here:

  1. Shouldn’t we deal with QUE elements one at a time, like we do with SLASH? and
  2. How should that be achieved, given that all wh-words, in-situ and not, are QUE-specified?

The ambiguity for (1) that I am talking about is this. Three trees, actually, not two. But I was talking mostly about the second and the third:

Screen Shot 2020-04-30 at 9.09.03 AM

I believe also that the whole point of something like the in situ question phrase is to “zero out” QUE… Under the current analysis. But as for filler-gap doing that, that seems hacky (and wrong, as the ambiguity shows). In the ERG, it was either not a hack or less of a hack because all lists there of at most of length one, so you could say that applying a filler-gap rule means either taking off one remaining element of QUE or identifying with the empty list. That’s my understanding anyway.

Also to clarify, if I do not say that filler-gap’s mother is QUE-empty (in a hacky way, ignoring what was on its daughters), then I end up with some sentences that contain more than one wh-word which are not QUE-empty at the end and so cannot pass the root conditions. And this is because, say, the head-complement rule appends QUE of its daughters, and we can have things like who says what.

For a verb-initial language, there would be no spurious ambiguity (extracted elements before the verb, in-situ elements after the verb). I don’t think reducing ambiguity is something you should be worrying about here – that seems like premature optimisation to me.

The ERG uses 0-1-dlists, so removing one element means the rest of the list must be empty. This is the same as the ERG’s treatment of SLASH that you wrote about in our HPSG paper. If you’re using append-lists and you remove one element (or any number of elements), you should pass the rest of the list up. That also means, if the mother should have an empty list, everything in the daughter’s list must be dealt with.

Of course, without an example of a QUE list with more than one element, this discussion seems hypothetical.

Under current analysis, any sentence with more than one wh-word, so, Who said what to whom? sort of thing. In said what to whom, QUE-list will have two elements on it.

Yes, I agree, what I’m saying is I discovered that this isn’t the case in my library right now, so, I am trying to figure out what is the right thing to do.

I am not sure it is important whether there is one or two elements on the QUE-list, but in both cases:

  1. Suppose we are analyzing Who knows who said what (to whom)?
  2. Said what has one element on its QUE-list (two if you add to whom)
  3. What should the QUE-value of Who said what to whom be? Who’s QUE-element will be dealt with by the filler-gap rule, presumably, but what about what’s QUE-element?

The fact that Who said what to whom can be embedded is not directly relevant here but it may become relevant later; it’s what eventually exposed my hack.

Furthermore, the first tree among those three is of course also problematic, and also exists because the filler-gap rule zeroes out QUE values.

Just a few remarks after our discussion about this:

  1. As @sweaglesw pointed out, this is very much connected with the fact that we aren’t (yet) representing the semantic connection between wh parameters and the questions they belong to (as discussed here).
  2. I don’t think it makes sense to talk in terms of a “unified” account of the NON-LOCAL features — they aren’t doing the same work and I don’t expect them to be handled the same way. That said, the head-filler and in-situ rules do both need to do something sensible with QUE, and it connects to questions of whether we want QUE values to be passed up beyond the current question.
  3. As you note above, I think your tree #1 in the annotated screen shot above is spurious, because it doesn’t constrain the SF value of the embedded question.

Perhaps I’m missing something, but why do we need a list here? There is no unbounded dependency. I would have thought the in-situ rule would deal with it.

Right, that’s exactly where we differ from Ginzburg and Sag’s analysis (which may or may not have been a wise move, ultimately). G&S have separate lexical entries for wh-words which are not fronted. They also have some additional pretty intricate “principles”, to make everything work for English. When we looked at that, we thought that analysis would not work for something like Russian (because of the additional principles). We also thought it would be nice to not posit additional lexical entries for wh-words. See part of the discussion here.

As for unbonunded dependeincies, in a language like Russian, both wh-arguments can be fronted, included when who said what to whom is an embedded question:

(1) Kto      chto     ty  tochno     znaesh,  chto skazal?
     who.NOM what.ACC 2SG definitely know.2SG that said?
"Who do you know for sure said what?" [rus]

When such wh-words are not fronted, we do need to license their position with regular head-subject and head-complement rules, and then, if none of them is fronted, with the in situ rule. Given that we also have free word order, we now have all this ambiguity, which I was able to keep more or less in check but now it turns out that maybe that was a bit too hacky.

If the issue is avoiding two lexical entries for each wh-word, how about a unary rule that converts from one type to the other?

I might still be missing something, but for this analysis of in-situ wh-words, it seems like the QUE list is accumulated and then thrown away.

Yes, you are right! And that bothers me, too, that’s exactly why I started this thread :).

The issue is also how to know when to apply the in situ phrase rule. If fronting is entirely optional, then all of the wh-words can stay in situ. So I’ve been operating under the assumption that something in the sentence must be QUE-nonempty, it is just that, for in situ, perhaps it matters less what exactly?

Supposing we use ICONS to connect wh parameters with events (as in that discussion), would this have any implications for the use of QUE?

I seem to remember that it was concluded that ICONS won’t work for this purpose after all; alas I do not remember why (and it looks like it was after Oslo, since it’s not in the notes).

Well, based on the discussion last summer, I think there are two conclusions: echo questions are a distraction, and this use of ICONS would affect truth conditions. (ICONS for anaphora would also affect truth conditions.)

In this thread, I was suggesting ICONS to be concrete. My question would also stand if we introduce a PARAMS list following Ginzburg and Sag: if we do something to connect wh parameters with clauses, does that have implications for QUE? Propagating QUE seems motivated for something like pied-piping, but seems less clear (to me) for in-situ wh-words, where accumulating a list seems like overkill. On the other hand, if accumulating the list is somehow helpful for connecting wh parameters, then that would be a good motivation.

1 Like

I’m the one responsible for repurposing QUE from pied-piping to wh in situ. I think it does make sense to collect things onto a list and discharge them at the level of structure that corresponds to their interpretation as wh parameters … if this is done with ICONS, then we’d be grabbing (say) the event of the head of the clause they combine with, building the ICONS, and taking that specific QUE element (or those specific elements) off of the QUE list.