when to annotate `compound` versus `obj`

[jonorthwash]

The documentation on compound states that compound isn't needed just because the meaning is lexicalised or idiomatic, giving example like make a decision.

However, the documentation on compound:lvc gives the example çile çektiler - literally "they endured suffering". This seems fairly compositional and non-idiomatic (unless you translate the verb more generally, e.g. "pulled suffering"). Why wouldn't an example like this be annotated with obj?

More generally, how can we tell when to use compound, and especially compound:lvc? The other two Turkish examples for compound:lvc seem reasonably like light verb constructions, since the verb et- really conveys very little lexical information in those examples. Is the criterion then about semantic content of the verb as compared to e.g. an object?

We're considering English examples like make money, make a decision, give permission, place an order (which feel like a continuum from more LVC-like to less LVC-like) as well as Kyrgyz examples like буюртма бер ("place an order", literally "give an order") as compared to уруксат бер ("give permission"), which differ in that the noun in the former one does not inflect and cannot have dependents, whereas the noun in the latter example can inflect and take dependents (*буюртмамды бердим, уруксатымды бердим). We are also contrasting these with idiomatic expressions consisting of a subject and a verb (e.g., башым айланды "I got dizzy", literally "my head spun").

The question is how to know whether/when to annotate such constructions literally (obj, nsubj) or as compounds (compound, compound:lvc).

[dan-zeman]

I am afraid that the borderline is not very well developed in the UD guidelines. My impression is (disclaimer: at the moment I don't have time to re-read the guidelines, so the impression may be wrong) that we essentially said that if people want to annotate LVCs as something special, they can go with compound:lvc, which should not be used in English, but there are languages where LVCs are much more prominent in the grammar and this might be a solution.

Obviously, this would not be / is not a good guideline when to use it and when to stick with the normal obj solution.

I completely agree with the part that says that semantic compositionality vs. idiomacity should not decide about UD syntactic relations. Which probably means that many such "compounds" are used wrongly and should be obj even in languages where LVCs are frequent (e.g. many Indo-Iranian languages). But as soon as morphology or syntax starts to behave differently from standard object-verb constructions, we can consider a different analysis, such as compound:lvc. So I guess that in your Kyrgyz examples, буюртма would be attached to бер as compound:lvc, while уруксат would be obj.

[coltekin]

I think the concept of light verb construction inherently relies on the verb being semantically light, and the weights of the verbs in these constructions are definitely graded. So, in general, it is really difficult to identify an intransitive :lvc without appealing to semantics. If the verb construction is transitive, then we can tell it with certainty. The Turkish example tercih et- 'to prefer' in compound:lvc is a good example - it takes and accusative object. Not marking it as compound would require having two objects (and, clearly this is not a double-object construction). On the other hand, analyzing tercih et- as compound:lvc and telefon et- 'to phone', which does not take an accusative object (its argument is dative, but no need to go into argument/adjunct discussion), not even as a compound would likely be confusing for the annotators. :lvc is a optional, language specific extension. So, it should be fine to appeal to semantics like other common examples (e.g., :tmod, :arg, :agt for obl). I think the main issue is determining compoundhood of them with certainty.

Returning to the compound definition, I find it is quite confusing. I think the guidelines wants to restrict the choice based on morphosyntactic criteria, but allows some "language specificity", and it is not clear - at least to me - what is exactly makes the positive examples in the documentation positive. From the examples, I get that noun-noun combinations without case marking should be compounds - at least for English. However, I also do not see 3 million dollar loan in the example having any level of non-compositional structure, which is given as the reason for the MWE relations including compounds. I think there is a need for a general rule/guideline here for determining what makes a compound more universally.

Perhaps more important for the current thread, the documentation is mainly focused on noun compounds, not saying much about verbs (except referring to language-specific documentation). It is clear that there are some cross-linguistically similar cases. So, it would be nice to put these together as much as possible, and include in the universal documentation.

[jonorthwash]

But as soon as morphology or syntax starts to behave differently from standard object-verb constructions, we can consider a different analysis, such as compound:lvc. So I guess that in your Kyrgyz examples, буюртма would be attached to бер as compound:lvc, while уруксат would be obj.

I wonder then about why this isn't the case in English, e.g.:

• place an order: placed my order, placed ten orders
• give an order: ?gave my order, ?gave ten orders (I've found examples of the latter, but they feel a little weird to me)

In a world where those '?'s indicate that such expressions are disallowed, then by your suggestion, wouldn't give an order have order as a compound:lvc dependent of give?

[nschneid]

IMO the strongest case for compound:lvc in English would be "have no idea" and similar idioms, because they can occur with a second NP whose function is difficult to classify if "idea" is an object: "I had no idea the time". UniversalDependencies/UD_English-EWT#444 That is, they are LVCs that trigger bizarre syntax. There are only a few expressions like this, though.

Otherwise I'm not sure there is much to be gained by applying compound:lvc in English as part of the syntax. There are all sorts of (semi)idiomatic multiword expressions that are beyond the scope of UD (but the PARSEME project annotates them as a separate layer). Those expressions may resist alternations that are characteristic of fully productive syntax, like modification and passivization (*the bucket was kicked is not a variant of the idiom kick the bucket)—but those alternation tests are used to describe the prototype of the syntactic function, and need not hold true of every single instance.

Taking a global perspective, I see your point that the definition of compound is a bit squishy, but this is probably necessary as different languages have different productive constructions that verge on creating complex lexical items, and are somewhere in between the rest of syntax and morphology. It is related to the issue of tokenization, where we are certainly not close to having a clear-cut universal standard.

Savary et al. 2023 discuss UD and MWEs, and suggest that the :lvc subtype could be removed from UD and relocated to an MWE annotation layer.

[jonorthwash]

There's quite a few types of N+V "compound" in Kyrgyz (and here I use the term to mean that the resulting meaning is not quite what one would expect if it were exactly semantically compositional—although this is maybe hard to make this call on in some instances, and it certainly isn't the deciding factor in UD).

Here's a brief typology I came up with some years ago, with some examples:

N with possessive morphology

N is 3rd person subject

Most of these are conjugated in 3rd person; English subject often corresponds to possessor of noun, overt with genitive case.

• ич+[pos] айлан- < ич ‘inside’ + айлан- ‘spin (v.i.)’ = "have an upset stomach"
• баш+[pos] айлан- < баш ‘head’ + айлан- ‘spin (v.i.)’ = "be/get dizzy / confused"
• жүрөк+[pos] айлан- < жүрөк ‘heart’ + айлан- ‘spin (v.i.)’ = "be/get nauseous / grossed out"
• көңүл+[pos] түш- < көңүл ‘feelings’ + түш- ‘fall’ = "be/get down"
• көңүл+[pos] көтөрүл- < көңүл ‘feelings’ + көтөрүл- ‘be lifted’ = "cheer up"
• көңүл+[pos] сууп кал- < көңүл ‘feelings’ + суу- ‘get cold’ + кал- ‘[perfect aspect]’ = "stop liking [a person]" (+ABL)
• кол+[pos] тий- < кол ‘hand’ + тий- ‘touch’ = "have free time"
• пайда+[pos] тий- < пайда ‘use’ + тий- ‘touch’ = "be useful" (+GEN - for the thing that is useful; verb is always 3rd person)
• суук тий- < суук ‘cold’ + тий- ‘touch’ = "catch a cold" (+DAT - for the one who got the cold; verb is always 3rd person)

N is definite object

seem to be either causatives of above or just compositional?

• баш+[pos]+[acc] айлант- < баш ‘head’ + айлант- ‘spin (v.t.)’ = "make dizzy / confuse"
• жүрөк+[pos]+[acc] айлант- < жүрөк ‘heart’ + айлант- ‘spin (v.t.)’ = "make nauseous / gross out"
• көңүл+[pos]+[acc] түшүр- < көңүл ‘feelings’ + түшүр- ‘make fall’ = "get [someone] down"
• көңүл+[pos]+[acc] көтөр- < көңүл ‘feelings’ + көтөр- ‘lift’ = "cheer [someone] up"
• сакал+[pos]+[acc] ал- < сакал ‘beard’ + ал- ‘take’ = "shave [one's] beard" (see also сакал ал-)
• уруксат+[pos]+[acc] бер- < уруксат ‘permission’ + бер- ‘give’ = "give [one's] permission" (see also уруксат бер-)
• боор+[pos]+[acc] эз- < боор ‘liver’ + эз- ‘mash’ = "crack [someone] up"

N is in other cases

• моюн+[pos]+[dat] ал- < моюн ‘neck’ + ал- ‘take’ = "accept/recognise/admit" (+ACC)

N without possessive morphology

"N" is mostly limited in use to compound

• кабыл ал- < кабыл ‘?’ + ал- ‘take’ = "accept" (+ACC)
• жөнгө сал- < жөн ‘~straight?’ + сал- ‘set / put in’ = "repair" (+ACC)

"N" is ideophone

• тарс эт- < тарс ‘bang’ + эт- ‘do’ = "go bang"
• чалп эт- < чалп ‘splash’ + эт- ‘do’ = "go splash"

N is indefinite object

• кол кой- < кол ‘hand’ + кой- ‘put’ = "sign" (+DAT)
• кол сал- < кол ‘hand’ + сал- ‘set / put in’ = "attack" (+DAT)
• сакал ал- < сакал ‘beard’ + ал- ‘take’ = "shave" (see also сакал+[pos]+[acc] ал-)
• баш тарт- < баш ‘head’ + тарт- ‘pull’ = "withdraw (from)" (+ABL)
• тамак ич- < тамак ‘food’ + ич- ‘drink’ = "eat / dine"
• баш ий- < баш ‘head’ + ий- ‘bend’ = "comply with / obey" (+DAT)
• баш кош- < баш ‘head’ + кош- ‘add’ = "become spouses"
• баш кой- < баш ‘head’ + кой- ‘put’ = "commit oneself (to something)" (+DAT)
• таасир эт- < таасир ‘influence’ + эт- ‘do’ = "affect" (+DAT)
• телефон чал- < телефон ‘telephone’ + чал- ‘~strike against’ = "call" (+DAT)
• буюртма бер- < буюртма ‘order’ + бер- ‘give’ = "place an order"
• уруксат бер- < уруксат ‘permission’ + бер- ‘give’ = "give permission" (see also уруксат+[pos]+[acc] бер-)

N is in other cases

• турмушка чык- < турмуш ‘lifestyle’ + dative + чык- ‘leave’ = "get married (for women)" (+DAT)
• камакка ал- < камак ‘captivity’ + dative + ал- ‘take’ = "arrest" (+ACC)
• жумушка ал- < жумуш ‘work’ + dative + ал- ‘take’ = "hire" (+ACC)
• башка кел- < баш ‘head’ + dative + кел- ‘come’ = "come up unexpectedly"
• жолго чык- < жол ‘way’ + dative + чык- ‘depart’ = "set off"
• жолго сал- < жол ‘way’ + dative + сал- ‘put in’ = "send off" (+ACC)
• жумушка сал- < жумуш ‘work’ + dative + сал- ‘put in’ = "set to work" (+ACC)
• сөзгө кел- < сөз ‘something said’ + dative + кел- ‘come’ = "come to an agreement"
• колдон кел- < кол ‘hand’ + possessive (person for whom its doable) + ablative + кел- ‘come’ = "be doable for someone" (+GEN & POSS)

Would we want Kyrgyz-specific guidelines for dealing with these, that involve compound relationships for some and not others?

[jonorthwash]

those alternation tests are used to describe the prototype of the syntactic function, and need not hold true of every single instance.

Got it, makes sense.

Taking a global perspective, I see your point that the definition of compound is a bit squishy, but this is probably necessary as different languages have different productive constructions that verge on creating complex lexical items, and are somewhere in between the rest of syntax and morphology. It is related to the issue of tokenization, where we are certainly not close to having a clear-cut universal standard.

It would be nice to know what the intention is / what the guidelines are, even if there's no exact definition currently.

[nschneid]

Potentially there could be Kyrgryz-specific criteria, or perhaps they could just be annotated as objects if there is not anything that makes them strikingly different morphosyntactically from regular (nonidiomatic) objects.

compound:lvc is a subtype that some languages have chosen to adopt, but there is nothing in the UD guidelines (that I know of) that actively recommends it for LVCs in new languages. The subtyping mechanism is mainly designed to give flexibility to treebank developers. A few subtypes are "semi-mandatory" (see here) but :lvc is not one of them. So, unless there is a pressing need to do something syntactically special with LVCs, I would probably not be in a hurry to use it.

[sylvainkahane]

I wonder then about why this isn't the case in English, e.g.:
place an order: placed my order, placed ten orders
give an order: ?gave my order, ?gave ten orders (I've found examples of the latter, but they feel a little weird to me)
In a world where those '?'s indicate that such expressions are disallowed, then by your suggestion, wouldn't give an order have order as a compound:lvc dependent of give?

Both are LVCs but with different senses of ORDER. I think that the contrast come from the properties of these two different senses (ORDER2 is less countable). From the syntactic point of view, both constructions are obj without any discussion.

I really agree with @dan-zeman and @nschneid that compound:lvc must be used when the LVC shows particular properties of cohesion, which suggest that it is no longer an object construction. The presence of another object, in a language where double object constructions are not regular constructions, is one of the possible justifications for introducing compound:lvc. More generally, compound should used for particular constructions that show a high level of syntactic cohesion and cannot easily be described with other relations.

In the French treebanks, we have introduced the relation obj:lvc for one reason: we consider that it is a regular obj relation but we wanted to indicate that in this case the complement of the (predicative) noun must be considered as the complement of a verb construction and not as a noun complement. Example:

Bill a besoin de sous. 'Bill needs money.', lit. Bill has need of money.
obj:lvc(a,besoin)
obl:arg(besoin,sous), rather than nmod(besoin,sous)

It is a case where MWE annotation and syntactic annotation intersect, at least with UD annotation scheme which distinguish obl and nmod.

[nschneid]

@sylvainkahane Interesting—I see the problem but must confess I find it counterintuitive to have obl:arg(besoin/NOUN, sous). From a basic UD perspective sous has to be either adnominal or adverbial, and this feels like it is trying to have it both ways (adnominal attachment but adverbial deprel). I think a noun is supposed to have obl dependents only if it stands for a clause, e.g. in a copular clause or when it is promoted due to ellipsis.

[jonorthwash]

Bill a besoin de sous. 'Bill needs money.', lit. Bill has need of money.
obj:lvc(a,besoin)
obl:arg(besoin,sous), rather than nmod(besoin,sous)

In my current thinking, I'm very sympathetic to something like this approach, except:

I think a noun is supposed to have obl dependents only if it stands for a clause, e.g. in a copular clause or when it is promoted due to ellipsis.

So I'd go with obl:arg(a,sous) instead of making the head of the object besoin. Then the verb is a besoin, and the oblique object, marked with de, is de sous.

I believe this is how I'd handle Yiddish ליב האָבן (lib hobn) "like", e.g. זײ האָבן ליב דעם הונט (zey hobn lib dem hunt), "they like the dog", with an accusative object in addition to ליב (lib). Or maybe this is okay to call compound:lvc? And זײ האָבן עס ניט ליב (zey hobn es nit lib) "they don't like it" would just have intervening obj and advmod, something like this?

1	זײ	_	PRON	_	_	2	nsubj	_	_
2	האָבן	_	VERB	_	_	0	root	_	_
3	עס	_	PRON	_	_	2	obj	_	_
4	ניט	_	ADV	_	_	2	advmod	_	_
5	ליב	_	PART	_	_	2	compound:lvc	_	_