Docs: Tough English sentence with "tough" adjective

[nschneid]

In the English syntax notes, the discussion of "tough" adjectives gives an example (112) for "This problem was hard for me to solve".

• What about "This was a hard problem for me to solve"?
• In (112), "me" is arguably the subject of "solve" rather than the modifier of "hard" as in (111)—though it may actually be subtly ambiguous. A clearer example is: "It is rare for students to graduate debt-free" (= "It is rare that students graduate debt-free").

[Note by CDM: I think the reference here is to the [v1 en specific syntax page](http://universaldependencies.org/docsv1/en/overview/specific-syntax.html); there isn't yet a v2 version.]

[nschneid]

@sebschu

[nschneid]

An even cleaner example that arose in our semantic annotation of prepositions: "I contracted for HDS to deliver the furniture." ("I contracted for HDS." cannot have this meaning unless it is construed as ellipsis.)

[dan-zeman]

In (112), I would maybe make nsubj(solve, me) a secondary dependency in the enhanced representation, but on the surface level I would keep nmod(hard, me).

In “This was a hard problem for me to solve”, I would do
nsubj(problem, This)
cop(problem, was)
det(problem, a)
amod(problem, hard)
nmod(hard/problem, me)
case(me, for)
xcomp(hard/problem, solve)

The subtle distinction with the two "hard/problem" heads depends on whether we understand it that it was “hard for me to solve” (and this entire phrase modifies the _problem_), or that it was a “problem for me to solve”, and BTW it was also _hard._ I believe that the first reading is more likely but, being a non-native speaker of English, I may be mistaken. And I do not think the semantic difference is too strong, so it may not be so important which of the two options wins.

[nschneid]

The subtle distinction with the two "hard/problem" heads depends on whether we understand it that it was “hard for me to solve” (and this entire phrase modifies the problem), or that it was a “problem for me to solve”, and BTW it was also hard. I believe that the first reading is more likely but, being a non-native speaker of English, I may be mistaken.

I agree: the more likely reading in my view is that it is a problem that is "hard for me to solve".

In (112), I would maybe make nsubj(solve, me) a secondary dependency in the enhanced representation, but on the surface level I would keep nmod(hard, me).

What about "I contracted for HDS to deliver the furniture."? nmod(contracted, HDS) would be a bit weird because I wouldn't say "I contracted for HDS"; it is the use with the infinitival that licenses "for". (Otherwise it would probably be "with": see PropBank examples of contract.02.)

[dan-zeman]

This sentence is outside of the limited English language model in my head :) so it is definitely for the English team to answer. I can see your point but if the weirdness means that we should attach "for HDS" to the infinitive, then I am not sure whether the borderline will be clear in other cases.

[nschneid]

@ngiordani, any comment? We're struggling over for-infinitivals in our data.

Another possibility that we've considered is to treat the for-PP as an argument of the adjective, and the embedded verb as an xcomp to indicate that it shares an argument:

This was a hard problem for me to solve
amod(problem, hard)
nmod(problem, me)
case(me, for)
xcomp(problem, solve)
mark(solve, to)

But in "I contracted for HDS to deliver the furniture.", it really does seem like "for" just marks the infinitival subject:

I contracted for HDS to deliver the furniture
ccomp(contracted, deliver)
nsubj(deliver, HDS)
case(HDS, for)
mark(deliver, to)

[ngiordani]

Sorry for the delayed response, @nschneid! The guidelines that we used in the EWT for tough-movement are given in this paper, pp. 2-3. By the guidelines, the analysis, updated for UD, would be as below... but keep reading, I suggest modifying this. Also, I'd ask if @tdozat has anything to add, since he's thought a lot about tough-movement.

PREVIOUS ANALYSIS
This was a hard problem for me to solve
nmod(hard, me)
case(me, for)
nsubj(hard, This)
amod(problem, hard)
ccomp(hard, solve)

(The paper explains why "for me" is considered an _nmod_, not an _nsubj_, and why _ccomp_, not _xcomp_, is used. I guess it's not that _nsubj_ is ruled out, but this is probably an _nmod_.)

After this paper, we decided not to use _ccomp_ under nonverbal predicates. In general, this decision is a little bit arbitrary and I won't say we are married to it, but in this case I think _advcl_ actually does make more sense, because clearly the clausal modifier is completely optional. So my recommendation would be:

PROBABLY BETTER ANALYSIS
This was a hard problem for me to solve
nmod(hard, me)
case(me, for)
nsubj(hard, This)
amod(problem, hard)
advcl(hard, solve)

In the other case I would argue for _mark_, and that is what we did in the EWT:

I contracted for HDS to deliver the furniture
ccomp(contracted, deliver)
nsubj(deliver, HDS)
mark(deliver, to)
mark(deliver, for)

I understand that the argument for _case_ would be that "for" licenses the overt subject, and that is certainly a good point.

One argument against is that we don't expect prepositioned subjects in other English constructions, so this would be the only place where we see an _nsubj_ with a _case_ dependent. (There is also the fact that "deliver" doesn't seem to select a prepositioned argument and that this PP seems to have a fixed position, but I admit these would be objections to labeled "HDS" with an _nmod_ more than to labeling "for" with _case_.

The other (related and maybe stronger) argument for _mark_ is that the appearance of "for" is tied to the infinitival form of the embedded predicate. That would be a strong argument for it to be understood as a dependent of that predicate.

[nschneid]

@ngiordani, thanks! It looks like your recommendations parallel the cleft examples (108) and (109) of the English syntax notes: (108) uses nmod/case, and (109) uses nsubj/mark. [Updated URL for link: [en v1 syntax specific constructions](http://universaldependencies.org/docsv1/en/overview/specific-syntax.html).]

[ngiordani]

That's true. Although, 112, which is basically the sentence we are talking about here, uses _xcomp_. I'm not sure how that happened, but it's wrong...

[nschneid]

It does seem to fit pretty well with the definition at xcomp. With the current analysis in (112), the xcomp makes it clear that "solve" should get a semantic argument from "problem" (in fact, it should get two!).

The paper linked by @ngiordani comments that it doesn't fit the classical definition of xcomp in LFG, but I wonder if the practical appeal of making it apparent that an argument should be inferred from a higher predicate should outweigh the historical boundaries of the term.

According to _CGEL_ (p. 542): "For the most part, complements in AdjP structure are optional elements: they qualify as complements by virtue of being licensed by the head rather than being obligatory." One of their examples of a complement is "He's [happy to leave it to you]." So I think advcl would be less appropriate.

[ngiordani]

I don't agree about _xcomp_. Open complements are defined to take their external argument from (the lowest argument of) the higher clause, which as you point out yourself, doesn't apply to the relation between "problem" and "solve". This isn't just a historical tie to LFG; it's a generalization that allows us to make inferences to identify arguments of predicates labeled _xcomp_. If you want to say that an _xcomp_ complement inherits some core argument from the higher clause, then we'd have a much harder time coming up with the rules for deciding which argument comes from where and making inferences. (Let me know if you think there is a simple way of designing such rules.) I don't think that's a good move, because as far as I can see, we would lose the robustness of the generalization.

With respect to _advcl_, well basically the idea is that _advcl_ is the clausal equivalent of _nmod_. So, being an _advcl_ doesn't necessarily mean the dependent is not an argument, it means it's not a core argument; the premise is that adjectives don't take core arguments (which I guess is pretty similar to saying they don't assign Case, in the language of GB, for example). So, even without committing to a decision about argument/adjunct, what you brought up is not incompatible with _advcl_. But this is a bit murky at the moment because there is an ongoing discussion about the whole idea of core arguments and how to define them in a way that makes sense crosslinguistically.

[nschneid]

If you want to say that an xcomp complement inherits some core argument from the higher clause, then we'd have a much harder time coming up with the rules for deciding which argument comes from where and making inferences. (Let me know if you think there is a simple way of designing such rules.)

In any case there would need to be a special rule to infer the semantic arguments for these for-infinitivals, right? I was thinking that calling it xcomp would make it clearer that _something_ needs to be inferred, and the rules would have to distinguish the different kinds of xcomps based on the syntactic environment. But I haven't tried to write such rules before, so perhaps there's too much ambiguity to do this deterministically from the Basic dependencies anyway.

[dan-zeman]

Tentatively closing. Feel free to reopen if unresolved issues remain in UD v2.

[nschneid]

I would expect to see "tough" constructions and for-infinitivals documented at complex clauses, but I don't, so reopening.

[dan-zeman]

@nschneid could you please propose the documentation, perhaps as a summary from the discussion in this thread?

[mcdm]

We had some suggestions for analyzing “tough” adjective constructions in our Depling paper 2013 (http://ufal.mff.cuni.cz/depling13/proceedings/pdf/W13-3721.pdf http://ufal.mff.cuni.cz/depling13/proceedings/pdf/W13-3721.pdf). Of course these can be revisited!

Marie

On May 1, 2018, at 2:33 PM, Dan Zeman [e-mail] wrote:

@nschneid https://github.com/nschneid could you please propose the documentation, perhaps as a summary from the discussion in this thread?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub https://github.com/UniversalDependencies/docs/issues/308#issuecomment-385750398, or mute the thread https://github.com/notifications/unsubscribe-auth/AHPFM8o5LqXFognfHmt7gEFoTyhNg71pks5tuKpugaJpZM4IpHEF.

[nschneid]

We had some suggestions for analyzing “tough” adjective constructions in our Depling paper 2013 http://ufal.mff.cuni.cz/depling13/proceedings/pdf/W13-3721.pdf). Of course these can be revisited!

Yes, @ngiordani made reference to the paper in this post, saying that there was a later decision not to use ccomp with nonverbal heads.

Is that decision still valid? I see a lot of adjective-headed ccomps in the data: "It is clear that...", "I am sure that", "I was able to", etc., quite apart from tough adjectives. Those certainly look to me like complement clauses, and in fact the new ccomp guidelines mention that the head should be a verb or adjective.

So should we go ahead with the 2013 paper analysis, namely:

This was a hard problem for me to solve

nmod(hard, me)
case(me, for)
nsubj(hard, This)
amod(problem, hard)
ccomp(hard, solve)

I contracted for HDS to deliver the furniture

ccomp(contracted, deliver)
nsubj(deliver, HDS)
mark(deliver, to)
mark(deliver, for)

?

[jnivre]

The status of adjectives as predicates (and the implications for labelling their dependents) has never been fully resolved in UD. My own feeling is that, as long as we use "amod" rather than "acl", we treat the adjective as a modifier word, not as a predicate, and should therefore not apply the core-oblique distinction to its dependents (which implies using "advcl" rather than "ccomp"). However, if we do use "ccomp", then it seems that we should also use "obl" (rather than "nmod") for "(for) me".

[nschneid]

I'm no syntactician, but there are clear parallels with verb complement clauses:

• Verb head:
  
  
  
  • They were informed of the problem.
  
  
  • They were informed that it was a problem.
  
  
• Adj head:
  
  
  
  • They were aware of the problem.
  
  
  • They were aware that it was a problem.
  
  

From an annotator's perspective, "that" + non-relative clause is a nice recognizable pattern; it would be convenient if it always signaled ccomp.

Can't we just say that predicative adjectives are predicates, whereas attributive adjectives (amod) are not?

if we do use "ccomp", then it seems that we should also use "obl" (rather than "nmod") for "(for) me".

Good point. The guidelines call for nmod only for modifiers of a noun/NP. In EWT the only ADJ-headed nmods are "many of" and "most of".

[dan-zeman]

if we do use "ccomp", then it seems that we should also use "obl" (rather than "nmod") for "(for) me".

Good point. The guidelines call for nmod only for modifiers of a noun/NP. In EWT the only ADJ-headed nmods are "many of" and "most of".

It could be read as an NP whose head noun has been elided, and the adjective _many/most_ has been promoted to the head position.

[jnivre]

You will never get ccomp for all that-clauses as long as we use acl in constructions like "the fact that ...". Note that UD doesn't have a complement-modifier distinction, it has a core-oblique distinction, and (at least according to the current guidelines), this only applies at the clause level. So the key issue here is whether the adjective is a clausal predicate or not, which it clearly is in "they were aware that ...". The fact that something switches from being ccomp to being advcl or acl when a construction is nominalised is a regular and expected feature of the annotation. Compare, for example:

She realised that ... ccomp
Her realisation that ... acl

[amir-zeldes]

I'm actually all for "realization ... that X" being ccomp. Treating it as acl removes the distinction between:

The realization that I was made aware of (true acl)
The realization that ice cream is delicious (a nominalized ccomp)

I understand the core/oblique distinction idea, but I'm not sure why we would want to not make the above distinction, which to me looks syntactically very clear (in one 'that' is substitutable by 'which', in the other not). This seems like a loss of information.

[jnivre]

That distinction is already made by having the first be acl:relcl. Using ccomp for dependents of nouns would require us to systematically distinguish complements from adjuncts in this position, which is even trickier than with verbs. I doubt that the benefits would outweigh the costs.

[amir-zeldes]

Isn't acl:relcl an optional/language specific distinction? I didn't understand acl without relcl to cover adnominal argument clauses... But more importantly, I think they're not really adverbial clauses. We can actually get all three:

The realization that I was made aware of (acl:relcl)
The realization that ice cream is delicious (ccomp)
The realization as it was described at the time (advcl)

Only the second one depends on the argument structure of realize/realization. The other two can be used with just about any noun. I thought the labels acl/acl:relcl, at least when they are distinguished as in English, just distinguish a canonical relative clause (finite, potentially with relativizer) from other types of adnominal adjunct clauses, e.g.:

A friend to go to the movies with (infinitival acl, but not relcl)

[jnivre]

Yes, subtypes are optional, so no one is forced to distinguish relative clauses from other adnominal clauses, but most treebanks do (I think). The crucial point is that, if the subtype is used, it distinguishes relative clauses from other adnominal clauses, not adnominal adjunct clauses, because UD does not acknowledge the argument-adjunct distinction. This is exactly parallel to the nominal case, where we use the same relation for obliques regardless of whether they are arguments or not:

I depend on you.
obl(depend, you)

I arrive on Tuesday.
obl(arrive, Tuesday)

And similarly inside noun phrases:

my dependence on you
nmod(dependence, you)

my arrival on Tuesday
nmod(arrival, Tuesday)

I am well aware that not everyone likes this way of cutting the cake, but it is built in at the core of the UD taxonomy, so making exceptions in special circumstances (like adnominal clauses) would lead to inconsistencies in the bigger picture.

[nschneid]

@jnivre, could you explain how the core/oblique distinction is tested for? For English verbs it seems straightforward because obliques equate to prepositional dependents of the verb. So I would expect that prepositional dependents of adjectives would be obliques as well, whereas that-clauses and infinitive-clauses that are dependents of adjectives would be core.

[nschneid]

(This page discusses core/oblique with respect to verbs, but doesn't mention adjective predicates.)

I see that overhauling the way we deal with adnominal clauses would be a major change. But I'm still confused on the issues regarding predicative adjectives. Let me see if I can get this right:

With a verb-headed clause, dependents include core nominals (subj/obj/iobj), oblique nominals (prepositionally marked), core clausal complements (ccomp), and adverbial clausal modifiers (advcl). That-clauses are always ccomp. Purpose clauses are always advcl. Other infinitive clauses may be ccomp, with the criterion "functions like an object of the verb, or adjective".

What does it mean to function like the object of an adjective? Pretty much the only obj(ADJ, *) relations in EWT are for worth, as in It is worth $10. So if "functions like" implies the ccomp clause alternates with an object for the same adjective, it doesn't make a whole lot of sense.

It seems to me that the simplest way to generalize the verbal-predicate guidelines to adjectival predicates is to say that kinds of dependent clauses that are ccomp in the former case are also ccomp in the latter case, and likewise for advcl. But maybe I'm missing something.

[dan-zeman]

It is actually questionable whether _$10_ is a core object of _worth_ (although I'm not sure what else it should be). It is obligatory and it is a bare nominal without preposition, so it does superficially resemble objects of verbs. But it cannot be promoted to subject via passivization.

[gossebouma]

While introducing a core/obl distinction for nouns does introduce additional work and decisions to be made, it is also true that our underlying treebank for Dutch makes this distinction, and it do not remember it as being an especially hard or controversial distinction. Most frequent nouns in our treebank with ccomp dependents are nouns like _fact, question, moment, risk, proof, hope, impression, danger,_ etc.

[jnivre]

I think many of the issues here revolve around two questions that have never been fully resolved in UD. The first is how to apply the core-oblique distinction to clauses. The original distinction from the typological literature applies only to nominal dependents, and many of the tests (passivisation, control, relativisation, etc.) are difficult or impossible to apply to clauses. It is therefore easy to reinterpret the core-oblique distinction for clauses as equivalent with the complement-adjunct distinction, which is not what we want. (@gossebouma I suspect that the distinction you refer to in your underlying annotation scheme is more like complement-adjunct clauses, rather than core-oblique clauses, isn't it?). The second question is how to treat adjectives in different uses. In the tripartite distinction between nominals, clauses and modifier words, adjectives are said to fall into the third category and are therefore expected to "allow some modification, but [...] not expand into the same rich structures as nominal phrases and predicates". But this is only clearly true of standard adnominal uses of adjectives (like "black cat"), while predicative uses clearly give rise to more complex structures. However, it is not clear that they can take objects like transitive verbs, for example. Going forward, I think we need a working group to sort out these and a bunch of other issues concerning the core-oblique distinction, given its centrality in UD. It would also be interesting to hear from @manning and @mcdm, given that many of the concepts and distinctions are inherited from Stanford Dependencies.

[dan-zeman]

The core/oblique distinction has been defined for verbs (rather than adjectives or adverbs). Verbs can be transitive but adjectives??? Probably not. Core objects are primarily observed with action verbs (while predicative adjectives tend to describe states). Then the notion is extended to some other verbs, and those verbs are called transitive, too.

I can imagine (or better: I'd be in favor of) extending the notion further to participles that are tagged ADJ (which is allowed in UD). But extending it to other adjectives and calling them transitive sounds weird to me.

If non-participial adjectives cannot have core arguments, then they cannot have ccomp (because UD defines it as core). Therefore, modifiers of adjectives are either obl (nominals), or advcl (clauses), or advmod (adverbs and adverb-like fixed expressions).

[nschneid]

Part of my confusion may be the terms "complement" in ccomp and "adverbial" in advcl. If those are not meant to reflect the argument/adjunct distinction, then perhaps different names are warranted (cobj and cobl?).

With respect to verbal clauses, if "alternates with obj" is really the criterion, then does that mean we'd have

• John said to Mary [that it was a problem]: ccomp, cf. John said 3 words to Mary.
• John informed Mary [that it was a problem]: advcl, cf. John informed Mary of the problem, *John informed Mary the problem, *John informed the problem to Mary
• John told Mary [that it was a problem]: in between—
  
  
  
  • John told Mary the problem (points to ccomp)
  
  
  • John told Mary of/about the problem (points to advcl)
  
  

This is starting to feel messy, plus I don't know if we want to force annotators to do a paraphrase test whenever they see a non-relative that-clause....

[amir-zeldes]

I agree with @nschneid that 'advcl' seems strange here, and with @gossebouma about annotation: we also annotated ccomp for nouns and adjectives in GUM using Stanford Dependencies (so we understood the SD guidelines to cover this), and it was not controversial among annotators. It is of course simple for us to remove the distinction in the conversion to UD, but it seems like a loss of genuinely interesting and syntactically motivated information.

Let me explain why I think adjectives and nouns can take core dependents, both as predicates and in some languages, as modifiers (when labeled amod). In English, there is a distinction between gerunds and deverbal nouns, studied extensively in theoretical syntax (e.g. Grimshaw 1990). We say:

The destruction of Rome by Nero (nominal, nmod)
Nero's destroying Rome (verbal, obj)

I fully agree with these labels: in the first example, Rome is simply prepositional, and we don't distinguish adjuncts and arguments (much like SD prep).

But in some languages, nouns and adjectives can take other nouns without adpositions, or even with adpositions explicitly coding for direct object. For example, Hebrew explicitly marks direct objects with the preposition 'et'. These are annotated as obj + case, and they appear with verbs and infinitives. But in Hebrew, pure verbal nouns (compatible with articles, adjective modifiers, etc.) are compatible with 'et':

shmiat-i et-ha-inyan - my hearing the matter.ACC (note shmia- 'hearing' is morphologically a noun for all purposes in Hebrew, distinct from the infinitive)

The same phenomenon can sometimes be found in Slavic languages, for example Church Slavonic:

о събрании съборъ - about assembly.LOC council.ACC 'about assembling the council' (again, deverbal noun, not infinitive)

And examples can be found for adjectives too. In German, an adjective like _ähnlich_ 'similar' takes dative dependents (we can argue whether this is obj or iobj of course):

... ist einem Tisch ähnlich - 'is [[a table]DAT similar]' (predicative)
[ein [einem beratenden Betriebswirt]DAT ähnlicher] Beruf] 'an [[an advising MBA]DAT similar]amod profession' (attributive) 

If we accept all of these nominal cases as 'core', then I think we should accept the much more common clausal ones as ccomp as well.

[dan-zeman]

@nschneid :

This is starting to feel messy, plus I don't know if we want to force annotators to do a paraphrase test whenever they see a non-relative that-clause....

Agreed. I'm afraid that our current guidelines are messy. I was just trying to show what I think they imply.

@amir-zeldes :

If we accept all of these nominal cases as 'core',...

If a modifier of an adjective uses the same coding strategy as a core argument of a verb, does it mean that it belongs to the same category as the core argument of the verb? I'm not so sure. As I understand it, one of the reasons why people want to treat the core arguments as something special is that the core arguments tend to be targeted by various grammatical rules (passivization is an example of such a rule). Will the similar-looking modifiers of adjectives be targeted by these rules too? Perhaps they will in languages where the borderline between verbs and adjectives is fuzzy. But I suspect that the examples and languages discussed so far in this thread will not qualify.

It looks like we need a version 3 of the UD guidelines where modifiers of adjectives and adverbs will be addressed separately from the core arguments of verbs, and from the modifiers of nouns. We may also want to redefine ccomp, or go with cobj and cobl (better parallelism with nsubj vs. csubj). But that is not going to happen tomorrow, probably not even within a year. So we need to find a solution that can still be claimed to be interpretation and better specification of v2 (rather than upgrade to v3), that is reasonably easy to apply, and that does not look too weird. If such a solution exists...

[dan-zeman]

BTW, the paper referenced above by @mcdm proposes ccomp for the relation between a tough adjective and the dependent clause. So this would be an argument in support of ccomp in UD, given that UD inherited ccomp from Stanford Dependencies. But then I think that we should abandon the claim that ccomp is a core argument.

[jnivre]

I agree with Dan's assessment. Sorting out all these issues is for v3. (But we should make it happen!) In the meantime, we need a workable set of guidelines that can be applied consistently across guidelines, and I think there is a lot to be said for the simple rule that only verbs take core arguments (including ccomp and xcomp). For borderline cases like gerunds, this means that we have to tag them as VERB if they do take core arguments in the same way as their finite counterparts but should use NOUN if they are true nominalizations (and there will always be tricky borderline cases).

[gossebouma]

In addition to @amir-zeldes observations about adjectives taking nominal dependents (without a preposition introducing the dependent), Dutch has several of those (I counted 30 lemma's in LassySmall), ie

De flashbacks zitten vol symboliek 'The flashbacks are packed-with symbolism'
Een mand vol levensmiddelen 'A basket filled-with food'
Conform de afspraken viel Polen onder de Russische invloedsfeer 'Conform the agreements, Polen fell under Russian influence'
Ik ben mijn portemonnee kwiit "I am my wallet lost" (I have lost my wallet)
Zij bleef het antwoord schuldig "She remained the answer missing" (She failed to give the answer")

There is some confusion in the annotation wrt elements like 'vol' which has POS adj but project to a phrasal PP (and the Alpino parser indeed just treats these as prepositions). Most other case occur where the adjective is used predicatively. In those cases, obj is maybe expected/allowed even under UD guidelines?

[amir-zeldes]

@gossebouma there is a very similar construction with 'full' in German, which is being grammaticalized from an adjective into a preposition - I coincidentally wrote a paper about this if you're interested: http://corpling.uis.georgetown.edu/amir/pdf/case_for_caseless.pdf

@dan-zeman and @jnivre - if this needs to wait for V3 that's fine, and I am happy to be on the working group for this, though I'm not sure why this is a major change: we're just talking about applying an existing label to some cases (I think it can remain ccomp, there's nothing wrong with that).

@dan-zeman - it's true that adjectives and nouns don't passivize, so we don't have that test, but that seems like an unfair bar: passivization is distinctly verbal, so naturally that test picks out verbs. If what you mean is 'expression of agent, nullification of agent, and optional oblique agent expression', actually all of these can be done by deverbal nouns:

The destruction of Rome (agent unexpressed, like passive)
The destruction of Rome by Nero (oblique optional agent, just like passive)
Nero's destruction of Rome (explicit agent in specifier position, just like active)

I agree that whether or not these are 'core' is a separate question, and of course the marking strategy here is different from the verbal case (except 'by'), and most similar to nmod/obl (including 'by'). So I'm happy with nmod(:poss) for these. However in English, that is not the case for clauses:

I decided that I want it
The decision that I want it

Both instances fulfill exactly the same kind of modification, with exactly the same marking strategy. So I don't understand why we cannot decide that this is still ccomp. Even if it requires a minor edit to the documentation, I think for me it qualifies as an 'interpretation and better specification of v2', not necessarily a major upgrade.

[kimgerdes]

In light-verb constructions a similar problem arises:

I took the decision to read that.
I decided to read that.

it seems preferable to have the same relation towards _to read that_ in the two examples: xcomp
If we decide to do that, we would also obtain a NOUN with an obl here:

J'ai besoin de toi. (fr.)
lit. I have need of you.
I need you.

This boils down to given preference to the syntactic function of the dependent rather than the POS of the governor in case these two constraints don't lead to the same relation.

We have 3 choices:

1. The category of the governor decides all: nouns always have acl and nmod
2. Predicative nouns can have obl and ccomp only if they are in a predicative position (ie. they are the dependent of a light verb).
3. Predicative nouns can always have obl and ccomp depending on their meaning. The predicativity of a noun can be tested precisely by the fact that it can appear in a light-verb position.

We (Sylvain and Kim) have a preference for the second or third solution.
An argument in favor of the second solution is the fact that some nouns do not behave in the same way when they are not in a predicative position:

J’ai besoin de toi.
J'ai besoin de partir, lit. I have need to leave
J’a besoin que Pierre m’aide, lit. I have need that Peter helps me.

but

?Mon besoin de toi est incessant, lit. My need of you is incessant.
?Le besoin de partir augmente, lit. The need to leave grows up.
?*Le besoin que Pierre m’aide est important, The need that Peter helps me is important.

And it is not so easy to decide whether a noun can occupy or not a predicative position. It seems that _fact_ cannot, even if we have pairs like:

I hate the fact that he leaves.
I hate that he leaves.

If we adopt the second solution and we restrict this to a predicative usage of nouns, it means that ccomp or obl on a noun would only be possible if this noun forms some sort of MWE with a verb and in particular a LVC, something which could optionally be annotated by a special sub-relation :lvc:

I took the decision to read that.
took -obj:lvc->decision.
decision -xcomp-> read

but

My decision to read that remains unchanged.
decision -acl-> read
I hate the fact that he leaves
hate -obj -> fact
fact -acl-> leaves

[dan-zeman]

@amir-zeldes :

... though I'm not sure why this is a major change: we're just talking about applying an existing label to some cases (I think it can remain ccomp, there's nothing wrong with that).

If the "some cases" are ADJ and ADV, then it is probably a small change because that area is now fuzzy anyway. If you mean NOUN, sorry, no. The line between dependents of nominals and dependents of clauses is quite clear in the current guidelines. Crossing it would be a major change no matter how few characters you need to describe it.

... it's true that adjectives and nouns don't passivize, so we don't have that test, but that seems like an unfair bar: passivization is distinctly verbal, ...

Isn't the notion of core arguments in language typology also distinctly verbal?

[amir-zeldes]

@dan-zeman - if you all say this is something that needs to wait for V3, then that's how it is, as consistency is certainly important. Our data is derived from an SD representation that makes the distinction, so it's easy enough to collapse/uncollapse.

However regarding the question: no, I don't think so, I think typologists generally accept deverbal derivations can preserve core arguments. It would surprise me if a majority of typologists would place importance on being a verb, but I guess we'd need a survey for that. There are many complex situations which would make forbidding this have side effects (the examples from @kimgerdes are a case in point).

But I also never understood UD to be taking a categorical position that only verbs can have typologically core arguments as a theoretical position. If so, why do we assign nsubj to nominal and adjectival predicates? Isn't nsubj (for typologists core dependent A/S) even more 'core' than obj?

I don't think having an obj is a matter of morphological category of the parent either. If we argue this way, then some languages could get rather odd paradigms. For example, in Japanese transitive verbs and desiderative verbs can take object (do X, want to do X), but desiderative forms are morphologically adjectival. There are several consequences to this: they take adjective-like tense inflection, and they can take an object marked as either nominative (-ga), or accusative (-o; I think this is an active fluctuation in Japanese). I just had a look and it seems at least in GSD, these things are marked as VERB + AUX and the object is governed as obj, so they are not noticeable to validation:

足を洗いたい
ashi o araitai
foot ACC wash-DESID
'want to wash feet'
obj(arai,ashi)

https://github.com/UniversalDependencies/UD_Japanese-GSD/blob/master/ja_gsd-ud-dev.conllu#L5161

I think it should be up to the Japanese team to decide if these are adjectives or verbs, but I think either way, they govern the object of what is being washed in the same way (or at least the team making the guidelines should have that option). For me part of the beauty of developing UD guidelines is exploring cross-linguistic variation. I'm not sure what purpose would be served by restricting obj to verbs (or verbs and adjectives).

[nschneid]

Speaking more generally, I think it's fair to say that linguistic structures evolve in messy ways: for example, perhaps English that-clauses started out exclusively as core clausal dependents of verbs, but then got reused with deverbal adjectives and nouns (pleased that, proposal that), then got extended to some non-deverbal adjectives and nouns (afraid that, the fact that). Given the morphosyntactic similarity of the that-clauses, why should the type of the relation be strictly constrained by the POS of the head? Are the relations better viewed as prototype-based categories, such that (for example) ccomp is used for constructions/marking strategies within a language that are prototypically for core clausal dependents of verbs, but may have been extended to other uses as well?

[dan-zeman]

Agreed that nsubj is a core argument and that it is used in nonverbal clauses. But my doubts are more about objects than subjects. I do not remember having seen any discussion of transitivity of non-verbs. In general it is better to refer to _predicates_ rather than to _verbs_ and both can be found in the literature. But so far I've seen nonverbal clauses typically defined in a different chapter than core arguments, and probably all examples of two-argument predicates that I've seen were verbal. So Amir's Japanese example is an interesting new discovery for me (although I think that Japanese is one of the languages where adjectives are very close to verbs, so it is unclear to me whether the observations are portable to Indo-European languages).

Yes, I too find discovering new language structures a wonderful by-product of the UD project. But we have to set some restrictions because one of the UD goals is to annotate same things same way. And then of course there is the counter-balance of not making all languages look the same, so if a restriction turns out not to work, it has to be modified.

[jnivre]

Let me try once again to explain why this would be a major change in the principles of the UD taxonomy. It is not really about verbs vs. nouns (and certainly not about postags). Nor is it primarily about whether nouns (or adjectives) can take core arguments, but about whether the same relations can be used both inside noun phrases and at the clause level.

First of all, the UD taxonomy is based on the assumption that the two most important types of structures found in natural languages are (i) clauses, where the head is a predicate (which may be a verb, noun, adjective or even adverb), and (ii) nominal phrases, where the head is a noun or noun-like word used referentially (in a wide sense) rather than predicatively. Second, we assume that dependents of (i) clausal predicates and (ii) nominal heads are labeled with distinct relations. This is why we have obl vs. nmod, advmod vs. amod, advcl vs. acl, and so on. (It was a clear deficiency of v1 that nmod had to do double duty in this respect.) Third, we assume that dependents of clausal predicates (but not of nominal heads) can be divided into core and non-core. This is why we have nsubj, obj and iobj vs. obl, csubj, ccomp and xcomp vs. advcl (but no corresponding oppositions for dependents of nominal heads, where nmod, amod and acl have to suffice for everything).

Therefore, adding a core-oblique distinction also for nominal heads would not only change the basic assumptions of the taxonomy, it would in principle also require a new set of labels in order to maintain the second principle above, that is, to use distinct labels for dependents of clausal and nominal heads, respectively. And this is definitely not something that can be added without a major version change.

Let me also add that some of the confusion regarding adjectives stem from the fact that, while the guidelines do recognise a third class of structures (besides clauses and nominal phrases), namely modifier words, the taxonomy does not contain distinct relations for dependents of such words. This is why advmod has to do double duty as (i) an adverbial modifier of a clausal predicate, and (ii) a modifier of a modifier word (adjective or adverb). (The latter function is what Bill Croft would call a "qualifier".) And for nominal dependents of modifier words, the current guidelines don't say clearly whether nmod or obl should be used, although I think the current practice in converging on obl, reserving nmod for nominal heads. Clarifying the guidelines for dependents of modifier words is definitely something that can (and should) be done already under v2 (in contrast to the issues discussed above).

Finally, I wish that @manning can find the time to comment on all these things soon, because he is to a very large extent the architect of this part of the UD taxonomy.

[amir-zeldes]

Thanks for that @jnivre , that does make the systematic issues clearer! I think that there are some discrepancies between current V2 practices and the division of relations above. For example, advmod is not just used as a modifier of modifier words in the NP domain, it is sometimes used as an adnominal modifier, and although I understand the system above, I'm not sure I find these wrong:

EWT example:
US troops **there** clashed with guerrillas - advmod(troop, there)

I think this is truly an adverb modifying a noun, and I don't think amod is better here. Some languages even have an adjective/adverb distinction like this, e.g. German _hier_ (adverb, 'here'), vs. _hiesig_ (adjective, 'here-y'). I also don't think another label is really required: it's really the same type of locative modifier as it is for verbs. I think an interrogative test makes it look compatible:

advmod
Troops where? Troops there. (cf. What kind of troops? Local troops.)
Went where? Went there.

Do we really need a separate label just because one is clausal and one isn't?

From the presentation above it seems there are two separate questions:

1. Do non-clausal heads have a core/non-core distinction (if this is taken to mean 'objects', I think the answer is sometimes, yes)
2. Should we consider re-using the same function labels across domains if the category is marked in the same way - so the original discussion boils down to, if there are adnominal core dependents, should 'that' clauses still be called ccomp

For the second question, it's clear that more splitting has no obvious structural disadvantages (you can always lump them again). But based on experience teaching dependencies in a lot of courses, I would say at least for English adding this distinction makes life harder for learning the scheme, and I don't see what it adds. 'Decision' and 'decide' can both take the same complement clause semantically and formally (same strategy), and I think it is definitely distinct from the relative one (substitutable by 'which'). And importantly, none of this gets us into the hot water of the argument/adjunct distinction, unless I'm missing something.

But I understand now that this is fully a V3 issue, so for now I'm thinking of using acl vs. acl:relcl as a place holder, if no one objects. If it needs to be something else I can do that too though, just let me know and I'll adjust the converter.

[jnivre]

Thanks, @amir-zeldes. A few comments:

• On advmod inside NPs: This is an old hot potato that was discussed intensely a few years back and probably never fully resolved. I agree that labelling adverbs "amod" would be misleading here, and we probably just have to acknowledge that "advmod" is a bit of a wildcard that can show up in different types of structures (just like "adverb" in traditional grammar).

• On splitting vs. lumping (issue 2): One of the advantages of the splitting strategy is that it facilitates conversion between different representations, for example, to facilitate parsing. Adding the "nmod/obl" distinction in v2 was very much motivated by the findings of Silveira and Manning (2015) where they tried to define a parsing representation for English.

• On placeholders: You can always add a subtype like "acl:ccomp" if you prefer (just like some people have added a subtype "obl:arg" as a way of preserving an argument-adjunct distinction from an underlying annotation scheme). Alternatively, you can put something in the MISC column.

[amir-zeldes]

OK, this makes sense. I think I prefer to avoid new subtypes in order to stay in sync with EWT, so I'll stick with acl vs. acl:relcl for now (@sebschu let me know if you have other preferences).

I think the reason I didn't see a problem with ccomp is that, unlike nmod/obl, there is no contrast between modifying an NP which happens to be a predicate, and modification of the predication (only case where an NP gets an obl modifier). If a predicate NP takes a ccomp, I think it's still modifying only the noun (maybe someone can think of a counter-example?).

I thought the main takeaway from Silveira & Manning 2015 was more or less 'no free lunch', at least for automatic parsing, because of the different behaviors of parser errors in the back-conversion. But since we're converting from gold SD I think we'll be OK either way for V2, and we can just change the rule in V3. If it seems acceptable for everyone to keep advmod as a single label, then maybe ccomp can also be used for the same type of expansion in V3, regardless of what is being expanded... I'm happy to join a working group on this.

[manning]

Sorry everyone (but especial @jnivre) that it has taken me so long to look at this! But I have just read through it all. I think I'll have to use as my example for discussion:

This was a very long thread for me to read

This thread started with adjective constructions and then veered off into noun constructions. I think the recent messages have made it clear that these should be treated separately, at least for practical reasons of what we could reasonably "clarify" in v2 versus making major changes to the adopted guidelines. I agree with and will follow this. This first post is on the status of adjective complements in general and then specifically returning to the _tough_-adjective questions. I might then write another post covering some thoughts on noun dependents.

@jnivre suggested in an earlier comment that there is sort of an inconsistency at the heart of what has been done whereby for nominal arguments of verbs we have clearly argued for a core vs. oblique (either oblique argument or adjunct) distinction, while for clausal dependents we have really been tending to make an argument vs. adjunct distinction. In particular, for these adjectival constructions with clausal dependents, we seem to be making an argument vs. adjunct distinction. I think he is essentially right about this. However, my pragmatic judgment is that this is the best thing to do and we should keep on doing it. 🙂 That is, I think it would be a mistake to replace advcl with cobl, since under item 5 of my desiderata for UD, "adverbial clause" is a broadly understood intuitive notion, whereas "clausal oblique" is a newly invented notion. However, it seems clearly wrong to be lumping what are arguably oblique arguments of predicates with adverbial clauses, and so I think they should stay as ccomp. (I would be open to renaming this as cobj since I think the notion of a "clausal object" would make sense to people, but in a way it is nice that "clausal complement" does in some sense capture that these clausal arguments may or may not be core arguments, and the term "complement clause" is reasonably widely known….)

With some differences or restrictions, in general when adjectives take arguments, they take them in the same manner as verbs. So I think the pragmatic choice for v2 is to say that dependents of adjectives can and should be marked in the same way as analogous dependents of verbs, as @nschneid suggested in one of the much earlier posts. I think this is fairly clear and uncontroversial when the adjective heads a clause/predicate like:

This thread was very long for me to read.

I'm still in favor of following the guidelines from the de Marneffe et al. 2013 paper, as adapted for UD v2. That is, we'd have:

det(thread, this)
nsubj(long, thread)
cop(long, was)
advmod(long, very)
root(ROOT, long)
mark(read, for)
nsubj(read, me)
mark(read, to)
ccomp(long, read)
punct(long, .)

Things get less clear/more controversial when the adjective is instead modifying a nominal (is attributive). However, my feeling is that if it takes dependents that we should still mark them in the same way as analogous dependents of verbs (or predicative uses of adjectives). So I would argue in favor of:

This was a very long thread for me to read.

nsubj(thread, this)
cop(long, was)
det(thread a)
advmod(long, very)
amod(thread, long)
root(ROOT, thread)
mark(read, for)
nsubj(read, me)
mark(read, to)
ccomp(long, read)
punct(long, .)

This is in essence saying that with an adjective (amod) there can still be a kind of reduced clause modifying the nominal head….

If you take this perspective, it seems like you should not a priori rule out adjectives taking other kinds of dependents that verbs take, even though it is a well known observation for English that in general adjectives do not take direct objects but rather obliques. There are some places where you might want to have an obj to an adjective. The example @amir-zeldes mentioned for Japanese is a good one (originally these desiderative adjectives took the "object of desire" as their subject, but there has been a widespread change in the last 30 years or so to be marking the "object of desire" with the object particle _o_; it seems reasonable now to call these an obj).

For English, having "transitive adjectives" is a very marginal property, but one that has been discussed in the theoretical literature. The clearest example traditionally is _worth_ in expressions like "That ring is worth a great deal." While arguably the amount phrase is not a true object, as was observed earlier in this thread, nevertheless the best relation to use for it in UD v2, it seems to me, is obj. A few years ago, there was a neat Language Log post by Geoff Pullum, citing an observation by Rodney Huddleston, whereby in modern financial English, adjectives like _underweight_ also allow apparent objects:

UBS says it is cautious on banks globally, and is underweight European and Japanese banks and neutral on U.S. banks. The broker is overweight Canadian and emerging market banks.

It seems to me that the obvious thing to do here in UD v2 would also be to regard this as an example of:

obj(underweight, banks)

[jnivre]

Thanks, @manning. So, in essence, you are making two proposals: (i) divide clauses into arguments and adjuncts, rather than core and oblique, and (ii) allow adjectives to take clause-level dependents in both attributive and predicative position.

I think (ii) is probably the right thing to do. It is more or less what we have converged on in practice, and it can definitely be seen as a specification of v2 and not a major change.

I am less convinced about (i) for two reasons. First, we run the risk of reopening the core-oblique vs. argument-adjunct distinction also for nominal dependents (why make the distinction for clauses but not for nominals?). Secondly, for some languages, this will create unnatural groupings of constructions. English is special in being very restrictive in allowing (nominal) clauses to occur with prepositions. But in Swedish (and I believe many other languages), we have regular pairs like:

Jag litar på dig.
Gloss: I trust on you
Translation: I trust (in) you

Jag litar på att du kommer.
Gloss: I trust on that you come
Translation: I trust that you come

Whatever is introduced by the preposition "på" is an argument in both cases (and the preposition is selected by the verb), but the nominal is definitely an oblique. If we are going to use ccomp rather than advcl in the second case (because it is an argument or complement clause) then we end up with the same verb taking "obl" and "ccomp" instead of either "obj" and "ccomp" or "obl" and "advcl" as we would expect.

[nschneid]

Jag litar på att du kommer.
Gloss: I trust on that you come
Translation: I trust that you come

Interesting! It seems we have two ways of looking at the head-dependent relationship: what the head licenses, vs. what grammatical encoding the dependent has. Perhaps canonically, these will match: verbs often alternate between licensing an accusative non-prepositional nominal and licensing a clause with complement marking (that in English, att in Swedish?). But some verbs, like litar, license an alternation between an oblique (prepositional nominal) and a clause with both på and att. As this construction with both markers seems to be language-specific, perhaps it deserves a language-specific label like ccomp:obl. Calling it "adverbial" seems counterintuitive given the common usage of "adverbial" to mean adjunct. And as @manning has said, "clausal oblique" is a nonstandard concept.

Also—practically speaking, as an annotator/someone who trains annotators, I am inclined to put higher priority on the grammatical encoding than on alternations that the head licenses, because it's easier to label what is there in front of you than to consider alternative sentences.

I understand the temptation to have complete symmetry in the classification that is made for clausal dependents as is made for nominal dependents. But to borrow a phrase from the Python community, I believe "practicality beats purity"—i.e. it's more important for UD users to be able to easily apply and understand the scheme in practice, using established concepts/terminology, than for the scheme to be perfectly elegant.

[adam-przepiorkowski]

The relation between the core/non-core and argument/adjunct distinctions is discussed in the “Arguments and Adjuncts in Universal Dependencies” paper (by myself and Agnieszka Patejuk, the developers of a new UD treebank of Polish making heavy use of enhanced dependencies) to be presented at COLING 2018. There we explicitly argue for the consistent application of the “no argument/adjunct” principle and for clausal obliques (introducing labels such as cobl and ccomp:obj; the paper was written and submitted in March, well before these labels appeared in this thread 🙂). One argument, similar to that of @jnivre, but stronger (I think), comes from coordination.

Take the attested:

“I asked him about his pregame rituals and whether he listened to heavy metal in the locker room.”

Is the coordinated dependent core or non-core? On the current setup, there is no coherent answer to this question: the PP conjunct is non-core, while the clausal conjunct is core. Hence, the fact that core/non-core distinction in the nominal domain does not correspond to the argument/adjunct distinction in the verbal domain leads to incoherence (IMHO).

A similar problem occurs in the case of xcomp, e.g. (the relevant bit is at the end of this lengthy attested example):

“Clare Hennessy, representing Merseyside police, said Campbell made two visits to the home of domestic abuse victim Miss A, during which he asked her for a kiss and to go on a date with him.”

Is the coordinated dependent “for a kiss and to go on a date with him” core or non-core? Again, there is no coherent answer: the PP conjunct is non-core, while the xcomp conjunct is core.

In the paper, where both examples (and more, also from Polish) are discussed, we suggest a way of applying the core/non-core distinction in the verbal domain, in a way that – we argue – is not only linguistically justified, but easy to implement (i.e., practical).

Apologies for the shameless plug – we sent the submitted version of that COLING paper to the recent posters in this thread a few days ago, but we'll be happy to send it to anybody else who is interested and might provide comments which would help us revise it.

Best,

Adam P.

[nschneid]

@adam-przepiorkowski, the coordination examples are interesting. To clarify: with

• He told me an idea and that he thought it was viable. — NP+clause coordination
• He told me about an idea and that he thought it was viable. — PP+clause coordination (*He told me about that he thought it was viable.) [I'm avoiding "ask" because both "ask whether" and "ask about whether" are valid, so the constituent structure is not obvious.]

Are you arguing that the first that-clause should be considered core because it's coordinated with an NP (direct object), whereas the second that-clause should be non-core because it's coordinated with a PP (oblique)? What is the basis for assuming that a core dependent can only be coordinated with another core dependent?

(Granted, the way UD handles coordination means that we aren't marking coreness on the second conjunct anyway....)

[nschneid]

Oh, I think I see—you're assuming that the coordinated material forms a phrase which has to be either core or non-core with respect to the verb? Because of the way UD handles coordination, the answer might be that the coreness of the phrase is determined by the first conjunct. Unless you're arguing that unlike phrase coordination is evidence that UD's approach to coordination should be overhauled. :) These cases seem different from canonical coordination anyway—for example, I don't know if the clause can come first (??He told me (that) he liked my outfit and about a friend with the same shirt).

[adam-przepiorkowski]

Yes, we are assuming that we want to avoid the situation where one conjunct is classified as core and the other as non-core, as then it is not clear how to treat the whole coordinated dependent. As to ‘inheriting’ from the first conjunct, Sag et al. 1985 (the classical “Coordination and How to Distinguish Categories”) give the following minimal pair:

• I didn't remember until it was too late John's inability to get along with Pat, and that he had no background in logic.
• I didn't remember until it was too late that John had no background in logic, and his inability to get along with Pat.

(In fact, they try to explain the latter away, as their theory does not predict it, but the fact remains that both are acceptable.)

It is true that sometimes only one word order is possible (when such a coordinated NP + CP phrase is a dependent of a preposition; relevant examples are again given in Sag et al. 1985), but in the usual case I think both orders are often fine when the NP is sufficiently heavy to follow the CP.

Of course we can call such examples ‘non-canonical’ and sweep them under the rug, but – at least in some languages – they are quite common. For example, in the largest Polish valency dictionary, Walenty, over 12% of valency schemata contain a position which has morphosyntactically different realisations which may be coordinated.

[amir-zeldes]

Hi all, thanks for the discussion! Just two thoughts from me:

1. about @jnivre 's example with 'litar på / litar på att': I think this is what we sign up for the moment we give up prepositional heads, and more specifically labels like Stanford's pcomp. The reason for this is not that the verb has two complementation structures, rather we can interpret it to say it takes the same complement in both cases: 'på'. Then it's the preposition that can take an NP or CP. I'd also like to point out that languages like German, where we have a correlate structure, are already following the ccomp analysis as follows:

ich zähle darauf, dass du kommst
I count thereupon, that you come
ccomp(zähle,kommst)
advmod(zähle,darauf)

We can argue about whether we think 'darauf' is an advmod, but in any case they are already in line with @manning 's suggestion.

1. I'm not sure what @adam-przepiorkowski 's example is showing: is it a problem that unlike coordination can occur? If I understood @nschneid correctly then I agree with him: it's the same as any case of coordination, where the left-most conjunct gets 'first dibs' in determining the relation. How is this different from the following:

Her coming and John's behavior annoy me (csubj + nsubj -> csubj(annoy,coming))
I want to stay home or at least in the city (advmod + obl -> advmod(stay, home))

If the point is that one conjunct is core and one non-core, then even prepositional dative coordination is problematic (I gave John a book and chocolates to Mary). I'm also not sure how this would mix with the idea that prepositionally marked clauses are necessarily oblique, what do we do when they're subjects?

For Kim to go is dangerous 
csubj?(dangerous,go)
advcl?(dangerous,go)

And of course these can be coordinated too:

Declaring war is dangerous and for Kim to go instead too.

[amir-zeldes]

I'm just seeing @adam-przepiorkowski 's last response: yes, exactly, so why enforce that both conjuncts must have the same coreness? I think that doesn't work out empirically, at least in corner cases.

I think the real way to solve it, if we have the resources, is using enhanced non-trees which mark the the second (or subsequent) conjuncts separately when needed.

[LarsAhrenberg]

I'd like to disagree with @amir-zeldes in his suggestion that taking the preposition as head in the constructions 'litar på / litar på att' would help. Although there are many problems with the UD analysis of prepositions as consistent case markers, in these cases the verb is critical for the selection of complements. There are other verbs that select prepositions that don't take clausal complements, such as 'hälsar (på)':

Hon hälsar på alla hon möter / She greets (on) everyone she sees
and others that accept a clausal complement but not an infinitival one (lita på; tvivla på / doubt) and vice versa. In all cases the preposition lacks independent meaning.
At the same time, I agree that a relation such as 'ccomp:obl' or 'cobl' could help make the parallelism between NP objects and Clausal complements clearer. In particular 'obj' and 'ccomp' (unmodified) would be clear correspondents.

[adam-przepiorkowski]

@nschneid and @amir-zeldes: Yes, an important assumption of the argument I gave is that coreness is a property of a whole dependent, so marking a part of it as core and another part as non-core is incoherent. I can see now that this assumption is not necessarily universally accepted. But if it is not, it is not clear (to me) that coreness means anything substantial in UD.

The difference between ‘unlike coreness’ and ‘unlike category‘ is clear. These days linguists do not assume that the category of the whole coordinate structure must be the same as the category of each conjunct. In fact, some believe that the category of the whole coordinate structure is the same as that of the first conjunct (e.g., Peterson 2004 in NLLT – within LFG, Zhang 2009 in her CUP book – within Minimalism), so UD may be on the right track here. But I think that – with the exception of the typologically and lexically very limited construction known as lexico-semantic or hybrid coordination, which receives rather different analyses than the usual coordination – nobody believes that two conjuncts bear different grammatical functions with respect to the governing head. For example, nobody – I think, although I am less sure now – would want to say that in “Pat remembered the appointment and that it was important to be on time” the conjunct “the appointment” is the direct object but the conjunct “that it was…” bears a different grammatical function. And nobody – I think – would want to say that the grammatical function of this coordinated dependent varies with the order of the conjuncts. But this is exactly what UD is saying at the moment in the case of “He asked her for a kiss and to go on a date with him” (etc.): here one conjunct (“for a kiss”) is oblique and the other (“to go on a date with him”) is core, so they are analysed as having different grammatical functions. Isn't that a conceptual problem?

[sylvainkahane]

@adam-przepiorkowski: UD is an annotation scheme, not a linguistic model. And UD relations are not grammatical functions. I hope that nobody beleives that nsubj and csubj correspond to two different functions. In French, an obj, an xcomp and a ccomp can very often fill the same grammatical function:

j'aime le chololat 'I like chocolate'
j'aime lire 'I like to read'
j'aime que tu lises 'I like you to read', lit. I like that you read

Maybe it would be better to have developped an annotation scheme based on grammatical functions, but it is definitely not the case of UD.

[jnivre]

I think it is too strong to say that UD is not based on grammatical functions. It is rather based on grammatical functions cross-classified with structural properties of the head of dependent. This is what is meant by a "mixed structural-functional system" in the guidelines (http://universaldependencies.org/u/overview/syntax.html).

[adam-przepiorkowski]

@sylvainkahane: Sure, but there is a connection between grammatical functions (in the linguistic sense) and UD's core/non-core distinction, isn't there? Doesn't the sameness of grammatical functions imply the sameness of the core/non-core status? If so, the reasoning is:
• conjuncts have the same grammatical function with respect to the governing head (common linguistic assumption)
• hence, they should all be core or all non-core.
If this reasoning does not go through, then I am not sure what the intensional content of the core/non-core distinction in UD is (as opposed to the extensional set of guidelines: “if it is a bare NP, it is core”, “If it is a PP, it is non-core”, etc.).

[amir-zeldes]

@LarsAhrenberg : you're right, certainly not all prepositions can take NP and CP complements freely, and this can be constrained by the verb. I'm also not advocating a return to pcomp, I'm just pointing out that in a pcomp analysis, the problem doesn't arise, since the verb governs the preposition as prep either way. I think ccomp:obl is an elegant solution, since it delineates the categories clearly, doesn't introduce a new main label for people who might not want to make the distinction or are growing weary of major changes, and is easy to automatically induce from plain ccomp (just check if there's an ADP child with mark).

@adam-przepiorkowski : I agree that like-function coordination is overwhelmingly more common, but in real data, unlike-coordination absolutely does occur, so it's not a hard constraint. The most prominent example is sylleptic zeugma, but those often sound constructed:

• their cakes make money and us fat (obj+xcomp)
• he made his apologies and for the door (obj+obl)

Real corpus examples are often more subtle, and I think this is perfectly natural, e.g. from GUM:

• He washed his neck and behind his ears (obj+obl: annotated as obj->conj->case in GUM)

I see how in the latter example one might want to say 'behind his ears' is somehow converted to an NP
(=[place] behind his ears), but I think if it weren't for the first conjunct, most people would annotate this as obl. Maybe the 'first conjunct is more important'-tendency is reflected in the preference to order these ?

[adam-przepiorkowski]

@amir-zeldes: I don't know about sylleptic zeugma :-), but in the case of “He washed his neck and behind his ears”, both conjuncts are best analysed as direct objects, I think. Such PP direct objects and subjects are extensively discussed in Ewa Jaworska's 1986 Journal of Linguistics paper, where she gives examples such as the following (any many more, from English and Polish, mainly locative and temporal), convincingly arguing for their direct object / subject status:

• They considered _after the holidays_ to be too late for a family gathering.
• _After the holidays_ was considered to be too late for a family gathering.

[amir-zeldes]

I fully agree that 'behind his ears' can fulfill the same thematic role as a theme phrase, but I think most UD annotators would probably label "he washed behind the ears" as obl + case, not obj + case, if only for consistency in not having obj + case in English... Because basic UD is just a dependency graph, these phrasal 'double duty' considerations have to come up short on one end or the other. In constituents you could do something like (NP (PP behind the ears)), but in pure dependencies we don't have that option.

Either way, I think the restriction on not having core + non-core coordination is an interesting and probably very strong tendency in language. But I don't think it can be a hard and fast annotation guideline, since natural language does contain intentional semantic zeugmas that defy this rule, and also the more mundane examples of the 'washing behind the ears' type.

[manning]

I agree that the Swedish example cited by @jnivre 6 days before this comment is quite compelling for needing clausal obliques. (This made me curious about what you do with these at the moment in the Swedish treebank – if I successfully managed with the Turku search tool and my non-existent Swedish, it seems that you label them as advcl. Right?) So maybe it would indeed be better to rename advcl to cobl after all, notwithstanding my earlier objection to that? (Though I am still somewhat sympathetic to @nschneid's initial comments that largely support my earlier position…. Thanks!) The other mentioned alternative of using ccomp:obl is possible in UD v2 and doesn't require giving up the familiar advcl (both good things!), but on the other hand, it is conceptually rather problematic, since we would then have a 3 way distinction for clausal dependents of clauses (core argument, oblique argument, adjunct), which we most deliberately avoided having for nominal dependents, so that must surely not be the perfect solution.

[jnivre]

Yes, we analyse them as advcl, because we essentially interpret advcl as cobl. I have a hard time making up my mind about this, because I do think advcl is a familiar and intuitive name for most people. On the other hand, if we really intend it to mean cobl, one could argue that the familiarity is misleading.

[jnivre]

I completely agree that we don't want a three-way distinction core-oblique-adjunct for clauses as long as we don't have it for nominals.

[nschneid]

This thread has gone on for long enough that I'm not sure where things stand anymore, but here's my current thinking:

I'm very happy with the core/oblique distinction as implemented for nominal dependents of verbs in English, because there's an easy rule: __if it has a preposition, it's oblique; otherwise it's core__. Rather than ask the annotators to think about the valency of the verb (which of the current arguments are optional, which arguments could be added or substituted), the test is based solely on the marking (preposition or not).

Can we define core/non-core for clausal dependents similarly? The test for core clausal dependents could be, __does the clause exhibit a marking strategy (specific to the language) that primarily/canonically alternates with core nominal dependents of verbs?__ For English, I think non-relative that-clauses are such a marking strategy—they canonically alternate with objects (I know [many facts about cats]/[that cats like boxes]) or subjects ([My cat]/[That cats like boxes] drives me crazy.) I think on the basis of this overt marking, it would be reasonable to extend this notion of coreness to that-clauses headed by adjectives, and perhaps even nouns. For Swedish, which can have a preposition plus att, perhaps the language-internal criteria would be different.

For to-infinitivals I think it would be helpful from a downstream usability perspective to distinguish I want to eat (complement) vs. I arrived to eat (purpose adjunct). But I concede that this is more of an argument/adjunct distinction, and wouldn't object to declaring all to-infinitivals core and using subtypes (as others have proposed) to distinguish argument/adjunct. Thus: I want to eat would be xcomp:arg and I arrived to eat (currently advcl) would become xcomp:adjunct.

Subordinating conjunctions seem to be the main way of marking non-core clausal dependents in English, so those would be advcl, cobl, or whatever label we want to use.

[nschneid]

__Logistical aside:__ Assuming that lots more discussion of these issues is needed, will enough people be attending NAACL or COLING that we can hold such a discussion in person? (I'll be at both.)

[jnivre]

Just a quick note to say that @dan-zeman is visiting us in Uppsala this week and we are hoping to put together some kind of proposal about this. I think it will be roughly consistent with what you suggest (although it is important to also think about how it generalises to other languages).

Unfortunately, I won't be at either NAACL or COLING this year, only ACL and EMNLP.

[amir-zeldes]

Yes, I also think either an in person meeting, or if need be maybe video chat could be useful here.

A few quick replies to the above:

• @nschneid - I think the Swedish att problem applies to English cases like "we agreed on us ordering pizza" too, which I think is clearly an argument

• Giving up xcomp : advcl - we currently maintain this distinction in the Stanford Dependencies of GUM so it's easy for us to keep it in UD too. I agree that it is absolutely an argument/adjunct distinction, but I think it's syntactically testable - we just have annotators try to insert 'in order to':

  • I arrived in order to eat
  • * I want in order to eat

I know we've been assuming that we don't want the argument/adjunct distinction on account of it being murky for obl in some cases, but it's starting to feel to me like we want to throw out the baby with the bathwater here. So I guess I'm thinking maybe we should go with optional :arg and :adjunct or something similar for all of these cases?