problem with annotation of "sadece" in UD_Turkish-BOUN

[jonorthwash]

In UD_Turkish-BOUN, the word sadece "only" is often treated as a dependent on a verb, even when it's serving to restrict something else.

For example:

# sent_id = news_1874
# text = Başka bir krize neden olan MGK toplantısı ise sadece 30 dakika sürmüştü.
1   Başka   başka   ADJ Adj _   3   amod    _   _
2   bir bir DET Indef   _   3   det _   _
3   krize   kriz    PRON    Pers    Case=Dat|Number=Sing|Person=3   4   obl _   _
4   neden   neden   NOUN    _   Case=Nom|Number=Sing|Person=3   7   acl _   _
5   olan    ol  VERB    Ptcp    Polarity=Pos|Tense=Pres|VerbForm=Part   4   compound:lvc    _   _
6   MGK MGK PROPN   Abr Case=Nom|Number=Sing|Person=3   7   nmod:poss   _   _
7   toplantısı  toplantı    NOUN    _   Case=Nom|Number=Sing|Number[psor]=Sing|Person=3|Person[psor]=3  12  nsubj   _   _
8   ise i   AUX Conj    Mood=Cnd|Number=Sing|Person=3|Polarity=Pos  7   cop _   _
9   sadece  sadece  ADV _   _   12  advmod  _   _
10  30  30  NUM NNum    Case=Nom|Number=Sing|NumType=Card|Person=3  11  nummod  _   _
11  dakika  dakika  NOUN    _   Case=Nom|Number=Sing|Person=3   12  obl:tmod    _   _
12-13   sürmüştü    _   _   _   _   _   _   _   SpaceAfter=No
12  sürmüş  sür VERB    Ptcp    Aspect=Imp|Number=Sing|Person=3|Polarity=Pos|VerbForm=Part  0   root    _   _
13  tü  y   AUX Zero    Aspect=Perf|Evident=Fh|Number=Sing|Person=3|Tense=Past  12  cop _   _
14  .   .   PUNCT   Stop    _   12  punct   _   SpacesAfter=\n

Here sadece 30 dakika sürmüştü means something like "continued for only 30 minutes". Surely sadece should be dependent on 30, and not the verb?

But that raises what kind of relation it should be. It's not quite an advmod because it can be a dependent on pretty much any part of speech (like only and just in English). This seems to me to be what advmod:emph is for?

[nschneid]

FTR, only and just are advmod in English, which doesn't use advmod:emph.

There are language-specific documentation pages for advmod:emph in Turkish and Tatar. I don't know if that reflects all treebanks in Turkish or just some of them.

[jonorthwash]

@iambusra, do you have thoughts on this?

[jonorthwash]

FTR, only and just are advmod in English, which doesn't use advmod:emph.

Does this pass validation, when you have an advmod dependent on a noun or a number or an adposition?

[nschneid]

Yep, the validator does not look at subtypes except a couple of the semi-mandatory ones, and advmod dependents of non-predicate non-modifier words is acknowledged in the guidelines: https://universaldependencies.org/u/dep/advmod.html

[jonorthwash]

I guess my question was more about advmod modifying anything. So, for example, what about modifer words with advmod dependencies—e.g., in only this time, would only be an advmod dependency of this (a determiner), or it dealt with differently?

[nschneid]

I would make it a dependent of time. There is a policy (that I only noticed recently) distinguishing "pure function words" that do not accept modifiers other than negation: these include case markers/adpositions and articles, and potentially other words. For English I would assume that demonstrative determiners should be considered pure function words. (As opposed to demonstrative pronouns, which can take relative clause modifiers, for instance.)

Looking at treebanks there are some quantifier determiners with modifiers like "yet another", "nearly all", "almost every". But I don't see any demonstrative determiners with modifiers.

[Stormur]

I guess my question was more about advmod modifying anything. So, for example, what about modifer words with advmod dependencies—e.g., in only this time, would only be an advmod dependency of this (a determiner), or it dealt with differently?

If something is an ADV, it is required to depend as advmod, also in this case. I think that using advmod:emph is a useful distinction, which also points to the fact that this is not really a "usual adverb". In general, if we annotate sadece or else as an ADV, there are no problems in making it depend as advmod, on whichever other part of speech, and this is probably the best solution (instead e.g. of identifying a DET sadece for noun phrases vs. an ADV sadece for predicates).

Since it is annotated as a function word, I also agree that its relation is with the head of the noun phrase, so dakika or time in current examples. But it might also be otherwise: in something like nearly all persons, I would propend towards advmod:emph(all,nearly). I do not think that determiners are "pure function words", which seems to refer to the multifarious class represented in UD by CCONJ/SCONJ/ADP/PART.

[jonorthwash]

, there are no problems in making it depend as advmod

The question wasn't about the deprel but the head.

in something like nearly all persons, I would propend towards advmod:emph(all,nearly).

This is my inclination as well, but @nschneid suggests this isn't attested in (English?) treebanks.

Is there at least consensus that this is a reasonable annotation of such structures?

Since it is annotated as a function word, I also agree that its relation is with the head of the noun phrase, so dakika or time in current examples.

I don't fully get why this should be different from nearly all persons.

[nschneid]

The pattern I am seeing in English treebanks is that prenominal "nearly all" is treated as a unit, but not prenominal "only this". This may have to do with the fact that "only" is a focusing modifier that applies to nominals in general, whereas "nearly" is a degree modifier of a quantity.

• Only this book needs to be returned.

• Only books that are overdue need to be returned.

• Only my aunt's casserole was not finished in its entirety.

• Nearly all books need to be returned.

• Nearly 30% of student borrowers will see some relief with this policy.

• *Nearly books that are overdue need to be returned.

But then we have to deal with the less common case where the modified quantity is not a det or nummod:

• Nearly the entire population of student borrowers will be helped by this policy.
• "nearly everyone"
• "nearly a quarter century"

The quoted ones are both treebank examples, and "nearly" modifies the head of the nominal ("everyone", "century"). So it is not a categorical rule that "nearly" never modifies the head nominal.

At the end of the day I don't think English UD has all the answers on determiners and determiner-related dependents in nominals, especially when it comes to various constructions for expressing quantity and measurement. So, take this precedent with a grain of salt.

[Stormur]

, there are no problems in making it depend as advmod

The question wasn't about the deprel but the head.

I was referring to this in the original post:

But that raises what kind of relation it should be. It's not quite an advmod because it can be a dependent on pretty much any part of speech (like only and just in English).

With regard to the head

Since it is annotated as a function word, I also agree that its relation is with the head of the noun phrase, so dakika or time in current examples.

I don't fully get why this should be different from nearly all persons.

I would also give an explanation like the one by @nschneid , but the more I think about it, the less obvious it is :thinking: Probably they interact with the other modifiers: degree modifiers can modify others, and focalisers can focus anything.

But it might as well be that I was making a blunder. Looking at these examples

• Nearly the entire population of student borrowers will be helped by this policy.
• "nearly everyone"
• "nearly a quarter century"

the position of nearly "outside" the determiner seems to point to it modifying the whole phrase, and so attaching to the head. Now we might talk about how much lexical or functional terms like entire (maybe PronType=Tot?) or quarter (a numeral) are, but the whole phrase is to be considered "scaled". This happens also for focalisers, as in only this book...

The reason for not being able of saying nearly books is that this expression is not quantified somehow. But I would no longer say that this means that if a quantifier is present, nearly attaches to it.

So personally I would correct my previous statement and lean towards attaching elements like nearly, only, etc. always to the head. In some cases, this might indirectly lead to reconsidering annotation of some elements (entire, quarter...)