VERB -?-> själv in Swedish and other adjectives used adverbially

[harisont]

In the material I am annotating I found a couple of instances of "...VERB själv", e.g. "att bestämma själv".

As I was unsure about the annotation of "själv", I checked the Swedish treebanks and found that both Talbanken and LinES use the UPOS tag ADJ, but there is a systematic difference when it comes to dependencies:

• Talbanken always uses VERB -amod-> själv
• LinES always uses VERB -advmod-> själv

The latter annotation fully aligns with my intuition, but I see that there are no instances of ADJ + advmod in Talbanken at all (vs. 68 in LinES). Is the underlying idea is that, if something is an adjective (or rather a pronoun, in this case?), it cannot be an adverbial modifier, and if so, how negotiable is that idea? In my view, dependency annotation is or should be more about what the words do (i.e. how the words are used) than about what they are, and the opposite goes for POS tagging.

[dan-zeman]

Is the underlying idea is that, if something is an adjective, it cannot be an adverbial modifier, and if so, how negotiable is that idea?

I don't know if there is such an idea at the level of Swedish-specific guidelines, but there is definitely no such limitation at the universal level. (Other languages, like German, may also need adjectives in adverb-like positions.)

If any restriction can be drawn from the universal guidelines, then I'm inclined to say that it is the opposite: You cannot attach an adjective as amod if the parent node is a verb. (But no such restriction is enforced by the validator, at least not yet.)

[harisont]

I fully agree! And yes, I was referring to the Swedish guidelines (well, actually, not even that: I was wondering if there's a principled reason why Talbanken uses amod.

[jnivre]

The lemma "själv" has at least three uses in Swedish:

• The first is as an adjectival modifier, which only occurs in definite NPs like "det är själva problemet". Here the relation is clearly "amod".
• The second is as a modifier of pronouns, as in "du borde vara dig själv", where Talbanken uses "acl", essentially analyzing it as a secondary predicate ("predikativt attribut" in traditional Swedish grammar). One can debate whether this is correct, but since it modifies a nominal head, it cannot be "advmod", and it doesn't go into the normal slot of an "amod".
• The third (which is the one under discussion) is as a modifier of a predicate, as in "bestämma själv". I agree that "amod" is unexpected here, so I think this is something that resulted from the conversion from the old annotation and never was noticed. One possibility is to use "advmod", but one argument against this is that it agrees with the nominal subject, which normally only adjectives do: "han bestämmer själv" vs. "de bestämmer själva". This is why the secondary predicate analysis was used in the first case, and the corresponding analysis here according to the guidelines would be "xcomp". I think this would be preferable, since the agreement shows that it is a predicate rather than an adverbial, but I am open to discussion.

[dan-zeman]

Secondary predication sounds like a reasonable analysis to me. Then the guidelines provide two options: xcomp or advcl (not advmod). Isn't själv an optional depictive? If it is, then the relation is advcl.

[jnivre]

Yes, you're right. It's optional, so it should be "advcl", not "xcomp", which goes well with the "acl" analysis when it has nominal attachment.

[Stormur]

I would also suggest advcl for what looks like some sort of floating qualifier/quantifier. More specifically, there is also the subtype advcl:pred to make it explicit that it is what is usually called a secondary predication.

Secondary predication sounds like a reasonable analysis to me. Then the guidelines provide two options: xcomp or advcl (not advmod). Isn't själv an optional depictive? If it is, then the relation is advcl.

I thought that the part about "optionality" and "depictive" was to be considered outdated (there were for sure many discussions in the past). Especially in the sense that I find extremely hard to describe what is usually labeled as xcomp as a "secondary predication"...

Is the underlying idea is that, if something is an adjective, it cannot be an adverbial modifier, and if so, how negotiable is that idea?

I don't know if there is such an idea at the level of Swedish-specific guidelines, but there is definitely no such limitation at the universal level. (Other languages, like German, may also need adjectives in adverb-like positions.)

If any restriction can be drawn from the universal guidelines, then I'm inclined to say that it is the opposite: You cannot attach an adjective as amod if the parent node is a verb. (But no such restriction is enforced by the validator, at least not yet.)

If that is the case, it is not clear at all from the guidelines. From them one could actually get the opposite impression: you read "An adverbial modifier of a word is a (non-clausal) adverb or adverbial phrase that serves to modify a predicate or a modifier word", and you are redirected to ADV. What possibly hides under "adverbial phrase" is not clear, and it sounds kind of weird to consider single adjectives as phrases: then why not also nominals (which are not so clearly excluded in the subsequent paragraph of the guidelines)? You know that I wrote an unreadable outsized paper exactly about this problem 😬 And it is still there!

[harisont]

Interesting, thank you all. advcl seems very reasonable to me, even though I hadn't thought of it at first.

[jnivre]

Great! Then I will put this on my list of things to fix in Talbanken for the next release.

[harisont]

The second is as a modifier of pronouns, as in "du borde vara dig själv", where Talbanken uses "acl", essentially analyzing it as a secondary predicate ("predikativt attribut" in traditional Swedish grammar). One can debate whether this is correct, but since it modifies a nominal head, it cannot be "advmod", and it doesn't go into the normal slot of an "amod".

By the way, I just noticed that LinES uses amod in this case. The same goes (in most treebanks) for similar constructions in Italian and Spanish ("sé stess*"/"sí mism*"), although I am not exactly sure how similar they are.

[LarsAhrenberg]

själv is a very versatile word and the description in SAOL very looooooong. If you regard DET/PRON/ADJ/ADV as a kind of continuous space you can easily find it all over the place and it seldom or never behaves as the prototypical words of these categories. For the time being, however, I stick to our analysis of själv as ADJ.

Normally it appears on its own without dependents. That's why I have annotated it as a modifier in LinES. (Unfortunately, nmod rather than amod in many instances, a rest from earlier days when själv was PRON.) I was surprised to see it analysed as a clause; clauses should normally have subjects and when minor, a place for a subjunction. Having read the guidelines on secondary predication I can understand why it should be advcl in UD, when governed by a verb, but I don't quite see the motivation for acl when it modifes a nominal as in chefen själv [the boss himself] or för sig själv [for/by herself/himself]. True, it is placed differently than normal ADJs but is that a reason to regard it as a clause when other clause features are lacking?

I note, for the record, that there are cases where it is tempting to use advmod rather advcl ie cases where the "depictive" analysis is less persuasive and similar alternatives are adverbs such as även or också (from Bellman's Fjäriln vingad)

Rörd och tjust av dessa blickar,
Själv den trumpne glättig går

[jnivre]

A clause in UD only requires a predicate, so as soon as we assume that a word is used predicatively, a clausal relation is in order. And there are many other cases where single adjectives are analyzed as clauses, notably in xcomp-constructions like "hon föreföll glad" and "hon gjorde honom glad". I therefore think it is natural to use "acl" when attached to a pronoun in analogy with the "advcl" analysis when it is attached to a verb. A possible counter-argument to this analysis is that it seems hard to insert anything between "sig" and "själv" in "sig själv", or move "själv" to a different position (which is possible for the predicative use with verbs).