Validator support for ExtPos in checking external deprels

[nschneid]

In English-EWT sentence answers-20111108103930AA7FPhc_ans-0007 there is a connective that is clearly supposed to be prepositional "due to" but the "to" is omitted.

The way "due to" is normally handled is as an ADJ+ADP fixed expression, functioning holistically like a preposition, which we indicate with ExtPos=ADP on the first word (#1037).

At present, the validator ignores external deprel checks on fixed heads. But in this sentence, the "to" is missing, so there is no overt fixed relation, and the validator is throwing an error that an ADJ cannot attach as case.

I think the correct validator behavior is to use the ExtPos if present for checking the deprel. I will temporarily change "due" from case to amod but hope to change it back in the future.

[amir-zeldes]

+1 - the correct deprel is case and not amod. An alternative is to directly tag "due" as a preposition in this context, but I like this suggestion better, since it's really just due an error ;)

[dan-zeman]

Implemented. In consequence, some treebanks have errors that were not reported before (because the treebanks use ExtPos and its value does not match the deprel):

• French-GSD ... 1
• French-Sequoia ... 2
• Portuguese-Bosque ... 12
• Portuguese-GSD ... 6

[arademaker]

Ok, I can fix the Portuguese GSD and Bosque.

[dan-zeman]

I am now going to gradually remove the exception for fixed expressions from the rel-upos-* tests, because these can be resolved with ExtPos in a more targeted manner. There will be thus more errors in more treebanks. All such treebanks will be put in the LEGACY status, giving their maintainers four years to fix the data.

[nschneid]

Thanks. Would it be worth adding a warning for ANY fixed head without ExtPos? Currently it doesn't flag that "according/VERB to/ADP" (fixed expression attaching as case) should have ExtPos but I think it's better with ExtPos=ADP.

[dan-zeman]

Thanks. Would it be worth adding a warning for ANY fixed head without ExtPos? Currently it doesn't flag that "according/VERB to/ADP" (fixed expression attaching as case) should have ExtPos but I think it's better with ExtPos=ADP.

I don't know. But according to will probably be flagged in the next round. I am modifying the tests one-by-one, and rel-upos-case had not been modified when you were asking but it has been modified now.

[nschneid]

"According" is tagged VERB so it can attach as case or mark per the deverbal connectives policy.

[dan-zeman]

"According" is tagged VERB so it can attach as case or mark per the deverbal connectives policy.

Shouldn't it now use ExtPos=ADP?

[nschneid]

There are two issues here: the general policy on VERBs as case/mark and the treatment of fixed expressions.

It looks like the validator change UniversalDependencies/tools@5d0d028 prohibits regarding, given, and such as VERBs attaching as case/mark. But the guidelines explicitly say it is OK and we never discussed repealing that in favor of ExtPos.

Assuming the single-word verbal connectives are allowed, my question was whether there should be a WARNING for any fixed expressions lacking ExtPos. I think that was the conclusion of the Core Group discussion.

[amir-zeldes]

My recollection matches Nathan's - using ExtPos for single word 'case' would be a new policy.

[dan-zeman]

Well, using ExtPos for single word case was the request with which Nathan started this thread – although that was a bit different because there the second word was omitted by mistake.

The change regarding VERBs can be reverted in the validator if desired. But the note that I had there from the time we discussed it in the core group was saying:

###!!! February 2022: Temporarily allow mark+VERB ("regarding"). In the future, it should be banned again
###!!! by default (and case+VERB too), but there should be a language-specific list of exceptions.

So now I thought that instead of implementing a language-specific list of exceptions, one could simply put ExtPos in the data.

[nschneid]

I found a note from Dec. 9, 2021: "Remove the categorical prohibition [on VERB/mark]; Dan will add a lexical list of exceptions (but it may take time)"

Perhaps ExtPos is a more economical solution than adding a lexical list. We should discuss in our next meeting. A concern is that we may be moving too fast in making ExtPos mandatory in some circumstances where it wasn't previously.