advcl:relcl for relative clause headed by an ADV

[nschneid]

A number of acl:relcl instances are headed by ADJs—these look OK as they are coerced into nominals ("the last", etc.).

But per the new guidelines, advcl:relcl should be used by ADV-headed free relatives ("where you work").

[amir-zeldes]

I'm not sure that guideline should apply universally to ADV-headed free relatives: in the "go" example, the "where" really serves an adverbial function. But if it's standing in for a direct object, for example, I would still expect acl:relcl, for example:

• I hate where/obj I live/acl:relcl

Here "where I live" is effectively a unary derivation converting the clause into an NP, so I think it's more intuitive to use acl:relcl, just as we give "where" the typically norminal label of obj (like you say, when the head is coerced into a nominal).

I agree that for predicative adjectives or true adverbs we should use advcl, for example:

• It was delicious/ADJ/root, which delighted/advcl:relcl all the guests.

[nschneid]

We already discussed this—I see the issue but the group decided it was simpler to use advcl:relcl based on a shallow look at the head being ADV. For cases like "go where/advmod you want", a full coercion analysis would have to explain that the where/ADV projects an NP with relative clause modfier which is then used adverbially (in which case the correct deprel would be obl:npmod, not advmod). UD doesn't give us the luxury of expressing all the levels of coercion, though, so we prefer adverb-style deprels where possible.

[amir-zeldes]

We already discussed this—I see the issue but the group decided it was simpler to use advcl:relcl based on a shallow look at the head being ADV

Really? I remember we discussed things like "John came which annoyed/advcl:relcl me", but if we had agreed to base the deprel purely on parent POS, then wouldn't that contradict what you propose for "the last" above?

I generally thin we should not be making deprels totally and automatically depend on parent POS - there is a reason those are two separate fields, and just as an ADJ can be nominalized but still tagged morphologically as ADJ, so can ADV be wrapped in a noun-y structure.

[nschneid]

Yes, we discussed in the 8/23 meeting. I agree we should not have a general rule that we ignore coercion or that all deprels are purely based on POS. But in the tricky case of adverb-headed free relatives, we decided it was simplest to take our cue from the POS.

[amir-zeldes]

Hm, this is fairly trivial to engineer for the WH cases... I don't really see the value, but it's no better or worse than having them all as acl:relcl (as long as they're all the same, we're doing injustice to one of the classes). What about non-WH ones? Or ones with a WH modifying an ADJ?

• Hapupu is a special place, and somewhere that you will want to tell/???:relcl others about
• I wonder how long it will take/???:relcl
• Sit anywhere you like/???:relcl (has a wh in it, but not xpos=WRB based on PTB)

[nschneid]

I think it's simplest to include all ADVs—so advcl:relcl for "somewhere" and "anywhere".

The second one is an interrogative complement clause ("wonder" doesn't take NP objects).

[amir-zeldes]

Whoops, nice catch (paradigm example of a complement clause, how embarrassing!)

But it's easy to find non-wonder-like ones too, where "how long" is a duration type object NP:

• I don’t know how long we’ll be (=I don't know the amount of time that it will take us)

For "somewhere" above, I feel like this heuristic basically makes the wrong call, since it's clearly a coordinate NP predicate (it is a special place and somewhere that...).

But like I wrote above, if we just make them all be advcl:relcl, it's not worse than making them all acl:relcl, and currently at least in GUM that distinction is not being made one way or the other.

[nschneid]

In reality I suspect "somewhere" has both adverb-like and noun-like properties. But "If a word is tagged as ADV, call its relative clause dependent advcl:relcl" is the most straightforward policy I can think of. :) Maybe some ADJs, too, but most of the EWT ADJs heading a relative clause were clear cases of nominal coercion ("the last"), so I left acl:relcl for those.

[amir-zeldes]

OK, I can live with this I guess.