MRS for extracted adjunct (adverb)

Here’s the MRS for “Where does Kim go?” from the ERG:

mrs

Here’s what I currently have in my development grammar for the Grammar Matrix library:

My question is, what about this very first relation from the ERG? The loc-nonsp thingy. What is it’s role? Do I want it for the Grammar Matrix?

There’s some documentation about loc_nonsp: http://moin.delph-in.net/ErgSemantics/ImplicitLocatives

The wiki says it’s relevant for ERG 1212, but it looks like it persists in ERG 1214 and 2018.

I don’t think it is exactly necessary (consider that “From where does Kim go?” does not have loc_nonsp because the location’s role is specific due to the “from”) depending on the analysis of locatives in your grammar, however the Russian MRS you have is disconnected. There is nothing linking _place_n with _go_v, so here you would need something, loc_nonsp-like or otherwise.

edit: in other words, loc_nonsp acts like a null-preposition in order to link the semantic fragments, and that may be useful when no other preposition exists in the sentence.

1 Like

Just to paraphrase, the ERG is effectively treating this as ‘to which place do you go’. Similarly, “I go there” will become ‘I go to that place’. In some languages (Japanese for example) this does not happen, “doko” is just ‘which place’ and the ‘to’ (ni) has to be explicit “doko-ni ikimasuka?”

1 Like

Thank you Michael and Francis.

This reminds me of so-called “D-linked” vs. non D-linked wh-phrases. Apparently you can extract things like “to which” more freely than “which”. E.g. in Russian you can probably extract multiple adjuncts (To which place in what time are you going?) so long as they are “D-linked” which I think just means there is a limited range of felicitous answers.

So I went back and started from just adverbs, not extracted adverbs. Trying first to get the correct MRS there.

Here’s one (for Ivan goes there); it is close but still does not have the TOP associated with the main verb’s LTOP.

16%20PM

I tried this, but this does not seem to have any effect (the LTOP part, that is):

my-head-adj-scop-phrase := head-adj-scop-phrase & 
 [ SYNSEM.LOCAL.CONT.HOOK [ INDEX #clause, LTOP #ltop ],
   HEAD-DTR.SYNSEM.LOCAL.CONT.HOOK [ CLAUSE-KEY #clause, LTOP #ltop ] ].

When I inspect the S node in the LKB, I do see the LTOP of the SYNSEM identified with LTOP of the head daughter, I think (??) but not in the MRS?

Anyway, I then tried the following:

my-head-adj-scop-phrase := head-adj-scop-phrase & 
 [ SYNSEM.LOCAL.CONT [ HOOK [ INDEX #clause, LTOP #harg ], 
 					   HCONS <! [ HARG #harg,
				                  LARG #larg ] !> ],
   HEAD-DTR.SYNSEM.LOCAL.CONT.HOOK [ CLAUSE-KEY #clause, LTOP #larg ] ].

This had a very interesting effect. I now have no parse for the sentence Ivan goes there but I have a unification result when I try the rule manually:
26%20PM

Help?!?

You want isect-mod-phrase, not scopal-mod-phrase.

Thank you, Guy. Could you explain why? I don’t think I really understand these types are originally for. The comments in matrix.tdl usually leave out definitions and so seem mostly arcane to me.

Also, still wondering in case someone knows: Why would it be that I have unification result but no parse?

Sorry to be late to this thread:

  1. I’m not sure that the LTOP is supposed to be the main verb’s LTOP there … we might be leaving that unconnected on purpose. I have a vague memory of something to with LTOP v. GTOP. Maybe something we should try to clarify at the Summit this summer?

  2. isect-mod-phrase is for intersective (aka non-scopal) modifiers. scopal-mod-phrase is for the scopal guys. The non-scopal ones are things like ordinary adjectives and PPs. The scopal ones are negation, things like probably, and also your clausal mods like because clauses. Something is scopal if it can participate in scope ambiguities with quantifiers.

  3. As for why you get a unification result but the edge isn’t in the chart, it may be one of the cases listed here: http://moin.delph-in.net/GeFaqUnifySurprise

Ah, bingo, I was just emailing you about this :). So, sounds like you actually like the second MRS? Except in the ERG, the TOP (and it’s not LTOP, by the way) is connected…

Yes, with two quibbles: The ARG0 of the unsp_adv_rel should be of type event, and I don’t understand why unsp_adv_rel instead of loc_nonsp. Is there some reason not to reuse the latter?

Oh, that’ was just a placeholder name! (The unsp_adv). Although, this way I can have just one relation for all adverbs? In the lexicon, time, reason, place will be specified; but the relation that is reserving the structure for the possible preposition can be one for all adverbs, maybe?

Unless you have a specific reason to differentiate, using the same one for all of them makes sense to me. “loc_nonsp” is something like “locative, non specific” or “locative, non specified” I think.

Alright.

I went with intersective-mod-lex for adverbs and, accordingly, with head-adj-int-phrase for my head-adjunct rule. So far so good, I have the same MRS as above (the better one, with the events properly linked).

One thing I am stuck on is the ARG0 of the extra adverbial relation (the one I called unsp_adv_rel for now). If I add the constraint: ARG0 event-relation, as below:

adverb-lex-item := intersective-mod-lex &
  [ SYNSEM [ LOCAL [ CAT [ VAL [ SUBJ < >,
                             SPR < >,
                             COMPS < > ],
                       HEAD adv &
                            [ MOD < [ LOCAL [ CAT [ MC +,
                                                    HEAD verb,
                                                    VAL [ SPR < >, COMPS < >, SUBJ < > ] ],      
                                              CONT.HOOK [ CLAUSE-KEY #clause, LTOP #ltop ] ] ] > ] ],
                   CONT [ RELS <! [ PRED "unsp_adv_rel", LBL #ltop, ARG0 event-relation, 
                                    ARG1 #clause, ARG2 #ind ], 
                                  [ PRED #pred, ARG0 #ind ] !>,
                          HOOK.LTOP #ltop ] ] ,
              LKEYS.KEYREL [ PRED #pred, ARG0 ref-ind & #ind, LBL #ltop ] ] ].

Then I get the error in adverb-lex-item: “Unification with constraint of ARG12-relation failed at path SYNSEM.LOCAL.CONT.RELS.LIST.FIRST”

I haven’t yet located a contradiction anywhere. Intersective-mod-lex is:

intersective-mod-lex := no-hcons-lex-item &
  [ SYNSEM [ LOCAL.CAT.HEAD.MOD < [ LOCAL intersective-mod &
					  [ CONT.HOOK.INDEX #ind ]] >,
	     LKEYS.KEYREL.ARG1 #ind ] ].

And no-hcons-lex-item is just:

no-hcons-lex-item := lex-item &
  [ SYNSEM.LOCAL.CONT.HCONS <! !> ].

What seems to be the problem?

You want [ ARG0 event ], not [ ARG0 event-relation ]. event-relation is a type that has [ ARG0 event ] as a constraint on it, so you could alternatively make that the type of the whole first relation on that RELS list.

1 Like

Alright, yay! Now I have this, and that’s using the intersective types, not scopal:

12%20PM

It’s still a bit mysterious why I can’t link the TOP… But at least the rest seems to be cleaned up :).

Almost — the _place_n_rel also needs a quantifier.

Ah, right.

So, I see that in the ERG, the quantifier is linked via hcons. Intersective-mod is a no-hcons item though which I think means I can’t use hcons.

Currently here’s what I did:

10%20PM

Should I use hcons instead? If so, probably means I need another type, not intersective-mod-lex.

The quantifier and the associated HCONS both come from the lexical entry for where. You can still use intersective-mod.

If I use a new type which is the same as intersective-mod-lex but does not inherit from no-hcons, then I can get the following:

55%20PM

I don’t really see how I can use the original intersective-mod-lex? Even if I push the HCONS constraints in the lexicon (which I still think I am misunderstanding you about!), those lexical entries would still be inheriting from adverb-lex and ultimately from intersective-mod and from no-hcons?