Subordinate clauses introduced by (relative) pronouns in Swedish

[AleksandrsBerdicevskis]

This is related to #1192, but I believe works best as a separate issue.

I've been looking at subordinate clauses in Swedish introduced by relative pronouns, and I think there are quite a few inconsistencies (or perhaps a pattern that I cannot find).
In most cases, the structure is as follows:

(1) ...ta ansvar för vad du gör
'take responsibility for what you do'

acl(ansvar,gör)
obj (gör,du)
mark(gör,för)

And this is the analysis that I would expect: the pronoun is labelled according to its role in the subordinate clause, and the subordinate clause depends directly on the main verb. If there is a preposition, it is a mark on the subordinate verb.

Examples: https://universal.grew.fr/?custom=696f9b4a50140

Quite often, however, the structure is different:

(2) ...hälften av vad en arbetare har i lön
'half of what a worker has as salary'

nmod(hälften,vad)
acl:relcl(vad,har)
case(vad,av)

In other words, the pronoun has a role in the main clause, and the subordinate clause depends on the pronoun. The preposition (in many such cases, though not always, there is a preposition) is a case on the pronoun.

I guess this analysis is reasonable, but I think the first one agrees much better with the UD spirit. Most importantly, I don't think the two can coexist!

Examples: https://universal.grew.fr/?custom=696f9b3d9df58

Questions, then:

1. Am I right in assuming that all cases in 2 are actually wrong and should be replaced by the analysis in 1?

(There may be some noise in the examples, and there are perhaps more that my query does not catch, but I hope they make my point clear.)

2. The pronoun sometimes has PronType=Rel and sometimes PronType=Int. Shouldn't it always be "Rel"?

3. I am less sure here, but shouldn't acl in (1) be acl:relcl?

[nschneid]

(2) looks like the analysis we have for free relatives in English (https://universaldependencies.org/en/dep/acl-relcl.html#free-relatives).

[amir-zeldes]

+1 for free relative analysis. I don't speak Swedish, but "vad du gör" in "för vad du gör" looks outwardly like an NP containing a clause to me, so the relation should not be acl from ansvar to gör. I think it should be treated as an NP (or PP, with case marking) headed by "vad", and similarly nmod(hälften,vad) in 2.

[jnivre]

There are a number of issues lurking here, but I think the main question is whether structures of this kind should be analyzed as subordinate clauses or as noun phrases that include a relative clause, or whether the two analyses are possibly both relevant for different cases. Let us start with the cases that do not involve prepositions or other markers:

(1) hon frågade [vem som hade tagit boken]
(2) hon frågade [vad de hade tagit]

In traditional Swedish grammar, the bracketed expressions in (1) and (2) would be called interrogative clauses, functioning as clausal complements of the verb. Hence, the appropriate relation in UD would seem to be ccomp. However, with respect to their internal structure, they look like they consist of an interrogative pronoun (vem, vad) followed by a relative clause, where the relative pronoun "som" is obligatory when the interrogative pronoun fills the subject rule but not when it fills the object role, which is the same as for ordinary relative clauses. This already raises the question of how to analyze the internal structure of the interrogative clauses, and I believe that the different Swedish treebanks don't agree on this. Talbanken treats "som" as nsubj and "vem" as dislocated in (1) and "vad" as obj in (2), but I believe that LinES has a different analysis (although I could be wrong).

Now, exactly the same kind of structures can occur after verbs that do not have interrogative meaning:

(3) hon berättade [vem som hade tagit boken]
(4) hon berättade [vad de hade tagit]

In traditional Swedish grammar, these would still be called interrogative clauses (I think), but some people don't like using that term when the sentence does not have interrogative meaning, and this may lead to the idea that they should be analyzed in analogy with the following similar sentences:

(5) hon träffade [mannen som hade tagit boken]
(6) hon hittade [boken som de hade tagit]

Here the bracketed expressions are clearly direct objects, with a head noun and a relative clause, and they look superficially similar to (3) and (4). I personally think the similarity is only superficial, and I think (3) and (4) should be analyzed like (1) and (2). One piece of evidence is pronominalization, where you have to use the neuter "det" in (3) and (4) but "honom" and "den" in (5) and (6). I don't want to rule out that there are genuinely ambiguous cases, although I can't think of any right now. And I remember similar cases being discussed for English.

I think the examples @AleksandrsBerdicevskis brings up are all instances of the same issue, although they appear as dependents of nouns with adpositions/markers, which complicates the picture.

[nschneid]

In traditional Swedish grammar, these would still be called interrogative clauses (I think)

(4) is 'She told me what they had taken', right? In the literature on English I have found, the subordinate clause is considered interrogative or 'indirect question' because there is sort of an implied question 'What had they taken?' which the matrix clause is about answering. Syntactic tests like sluicing (She told me what they had taken and why) distinguish this from the core free relatives (She ate what I had cooked (*and why)).

[AleksandrsBerdicevskis]

@jnivre Do I understand it correctly that you would analyze your (1-4) and my (1-2) in the same way, namely, as in my (1)? Mutatis mutandis, of course, wrt to the type of subordinate clause may, the syntactic role of the pronoun etc.

Not all my examples contain dependents of nouns, consider, for instance:
Pappa, mamma och barn bildar vad man brukar kalla för kärnfamiljen.
Talbanken: obj (bildar, vad); acl:relcl (vad, brukar).
Again, I would go for: ccomp(bildar, brukar), obj (kalla, vad).

@nschneid @amir-zeldes I agree that these examples are free relatives (as opposed to bound relatives), but I do not really understand why they have to be analyzed as having the pronoun as the head (though I guess that's how Huddleston and Pullum would analyze them).

In informal terms, I'd say that the pronoun has a double role here (regardless of whether it's a free relative or not): it is (or at least it can be viewed) both as part of the main clause and the subordinate one. I think different traditions have preference for different analyses (roughly as my 1 or as my 2, or one of those depending on the relative type) and all three solutions are defensible.

[nschneid]

Huddleston and Pullum analyzed it in a constituency framework with a device called 'fusion of functions' that allows for the double role that you describe. More at: https://doi.org/10.1017/S002222670700477X

In converting to basic dependencies, we have to lose one of these functions. I think the logic is that the higher one in the tree wins: the whole thing functions like a nominal phrase, and the external dependency coming into the WH word reflects that. The other would be available in enhanced dependencies.

[jnivre]

@AleksandrsBerdicevskis I realize that I only addressed part of your questions, and I also needed time to reflect and remind myself of earlier discussions with @LarsAhrenberg. There seem to be at least three issues involved here:

1. Should all the expressions that start with an interrogative pronoun be analyzed as clauses, or should some of them (specifically free relatives) be analyzed as nominals?
2. If some of them should be analyzed as nominals, where do we draw the line?
3. What is the internal structure of these expressions when analyzed as clauses and/or nominals?

I will leave number 3 aside for now and concentrate on 1 and 2. One argument for distinguishing two analyses, one clausal and one phrasal, in Swedish comes from the observation that "vad" can sometimes but not always be paraphrased as "det som":

(1) hon frågade vad de hade tagit
(2) *hon frågade det som de hade tagit
(3) hälften av vad en arbetare har i lön
(4) hälften av det som en arbetare har i lön

I think everyone (?) would agree that (4) should be analyzed as containing a (pro)nominal object "det" with a relative clause attached to it, and it is therefore tempting to say that the analysis should be the same for (3) except that "vad" merges the function of "det" and "som" into a single word, which is typical of free relatives. Because of this "fusion", we are forced to prioritize one function over the other, and treating it primarily as a nominal (the "det" part, rather than the "som" part) facilitates the analysis of its external role (that is, the relation to the rest of the sentence) and suppressing its internal function is no worse than for ordinary relative clauses without a relative pronoun ("lönen de fick" vs. "lönen som de fick"). I think this is the main argument for the analysis of free relatives in English, and am now inclined to say that we should adopt the same analysis for Swedish.

If we accept this, then the next question is whether we can answer question 2 above simply by relying on the paraphrase test. Here I am less sure, because paraphrases can be tricky, so I would like to hear what others think.

[LarsAhrenberg]

In my review of these constructions in Swedish_LinES I have tried to follow the difference @jnivre explained above, (but inconsistencies may of course still remain.) So, when you can replace the wh-word with a corresponding non-wh pronoun or adverb, then treat it as a free relative. In many cases you can also rely on the matrix verb: if it generally accepts objects or obliques but not clauses, as äta/eat or träffa/meet then it is a free relative. Or else, when in a nominal context as in @jnivre's examples (3) and (4). I doubt however that we need to introduce the notion of "fusion" here as I think som can be present or absent in both context. For me, the following are fine:

(3') hälften av vad som en arbetare har i lön
(4') hälften av det en arbetare har i lön

Now I agree with both @jnivre and others that paraphrasing in this way is not always easy to apply as there are ambiguities (examples at https://universaldependencies.org/en/dep/acl-relcl.html#free-relatives) as well as uncertainty in judgement. Anyway, my answer to Joakim's question (2) is that we can find more than one criterion on each category to separate free relatives from subordinate interrogative clauses and, as we cannot in UD make the distinction in terms of clause features, we may make the distinction as suggested for English with different internal structures.

[AleksandrsBerdicevskis]

As @LarsAhrenberg correctly points out, we need to think not only about pronouns, but also about (interrogative/relative) adverbs. As far as I can see, currently almost all constructions in all Swedish treebanks introduced by such adverbs receive the clausal treatment.

@nschneid: thanks for the clarifications! How exactly would enhanced dependencies capture the internal role? Will it be something like that:
half of what a worker has as salary

Basic:
nmod(half, what)
acl:relcl(what,has)

Enhanced:
obj(has,what)?

[jnivre]

Thanks, @LarsAhrenberg, for refreshing my memory about our discussions. I agree that we should try to follow the distinction you propose, relying on whatever criteria are available for different special cases. This will also (as far as I understand) make the analysis of these cases parallel in Swedish and English. I will put on my todo-list to review these constructions in Talbanken and PUD for the next release.

For me this just leaves the question of how to annotated the internal structure of interrogative clauses, specifically the "vem som" case when the question concerns the grammatical subject. The analysis that I have so far assumed for Talbanken and PUD is based on the assumption that this is a doubling of the subject and treats "som" as nsubj and "vem" as dislocated (in analogy with "Lars han bor i Linköping"). Judging from the few examples I could find, it seems that LinES adopts the same analysis, but maybe this is worth discussing given that the interrogative pronoun is assigned the core argument relation in cases where there is no "som". A potential alternative would be to treat "vem" as nsubj and "som" as expl, treating it instead as a case of "clitic doubling". I personally think this is less natural, though.

[LarsAhrenberg]

@jnivre your second proposal is very interesting and I guess you would not propose it unless you see some value in it. For myself I haven't thought about it before but I can see that som here have some typical properties of an expletive. Otherwise I see no problem in assigning vem an argument role in interrogative clauses such as vem du ska träffa / vem du har talat med and a dislocated-relation if the som is present. Or should I?

[nschneid]

@nschneid: thanks for the clarifications! How exactly would enhanced dependencies capture the internal role? Will it be something like that: half of what a worker has as salary

Basic:
nmod(half, what)
acl:relcl(what, has)

Enhanced:
obj(has,what)?

Enhanced:
nmod(half, what) (same as basic)
acl:relcl(what, has) (same as basic)
obj(has, what) (additional dependency showing the function within the relative clause)

Note that the enhanced graph does not need to be a tree. "what" has two incoming edges to reflect its double role.

[jnivre]

@LarsAhrenberg To be honest I thought that you had a different analysis of "vem som" in LinES, but when I saw that you didn't I felt I had to think about what could be an alternative. But I am not seriously proposing the expletive analysis, so unless someone else wants to promote it I think we can just agree to leave things as they are (with respect to this part of the analysis).

[sylvainkahane]

We have exactly the same problem in French. French has a +human interrogative pronoun qui and +inanimate interrogative pronoun quoi:

je me demande qui appeller 'I wonder who to call'
je me demande quoi acheter 'I wonder what to buy'

But quoi can no longer be used in non-infinite interrogative clauses and has been replaced by ce que, a demonstrative pronoun and a relativizer:

je me demande qui tu as appelé 'I wonder who you called'
je me demande ce que tu as acheté 'I wonder what you bought', lit. I wonder that that you bought

I don't have a perfect solution to propose. It is just an additional argument of the fact that interrogative pronouns have two syntactic roles. A solution would be to consider that the interrogative pronoun is the head of the interrogative clause. It would also explain the sluicing (You called someone and I wonder who). From the semantic point of view, the argument of the verb wonder is the interrogative pronoun:

I wonder who you called = I wonder the list of _x_s, _x_ being someone you called

[AleksandrsBerdicevskis]

@jnivre @LarsAhrenberg It looks like some kind of consensus has emerged for Swedish, can you check whether I understand it correctly?

Interrogative clauses:

(1) hon frågar vad de tar: ccomp(frågar,tagr), obj(tar,vad).

(2) hon frågar vad som händer: ccomp(frågar,händer), nsubj(händer,som), dislocated(händer,vad).
som is a pronoun (See #1149).

PronType for vad should be Int, and I am not sure about som. Rel?

Relative clauses (bound):

(3) hälften av det som en arbetare har i lön: acl:relcl(det,har), obj(har,som).
PronType for som: Rel.

Relative clauses (free/fused):

(4) hälften av vad en arbetare har i lön: nmod(hälften,vad), case(vad,av), acl:relcl(vad,har). Enhanced dependencies would also have obj(har,vad).
Currently, many such cases have acl and not acl:relcl, but it should be acl:relcl, right?
I am not sure about PronType for vad. Would it change if it's vad som in the same context, or det, or det som?

What happens with adverbs? Do we make a similar distinction between, say.
(5) Jag vet vart du går
and
(6) Gå vart du vill?