DK: usage of unknown dependency obl:loc

[TomazErjavec]

In preparation for 3.1 we are now using the common UD-SYN taxonomy. But using it with the DK corpus gives many errors like:

ERROR ParlaMint-DK_2017-05-18-20161-M99.ana: ERROR: Can't find local id for link/@ana="ud-syn:obl_loc"
...
ERROR ParlaMint-DK_2014-10-07-20141-M1.ana: ERROR: Can't find local id for link/@ana="ud-syn:obl_loc"
...

obl:loc does not seem to be a legal UD syntactic relations, so what to do with this now? We probably can't leave it as it is, as it doesn't mean anything. Change it to simple obl? Ideas welcome!

[matyaskopp]

For Danish, an old 2.5 model is used. In Universal Dependencies 2.5 – Danish – DDT was this relation used: http://hdl.handle.net/11346/PMLTQ-0RGX (612 occurences) But the taxonomy corresponds to the current version of UD (current documentation, to be precise).

note for @matyaskopp: this generates query with all ids with obl:loc that can be run in different version of ud to see changes http://hdl.handle.net/11346/PMLTQ-COKG

In 2.12 they have been replaced with: http://hdl.handle.net/11346/PMLTQ-8FW4 relation	occurences
obl:lmod	48
obl	1
case	3
advmod:lmod	560

---

So, the question is whether we want to support old undocumented language-specific relations. It is in some old statistics: https://github.com/UniversalDependencies/docs/blob/97694404898cc696842234a1ebabb888c448f09b/_includes/stats/da/dep/obl-loc.md But in fact, it has never been documented, The Danish language has not ever documented specific relations in the whole history:

git clone git@github.com:UniversalDependencies/docs.git Scripts/UD-docs
git -C Scripts/UD-docs checkout pages-source
git -C Scripts/UD-docs log --all --full-history -- "_da/dep/*"

Returns an empty result.

[TomazErjavec]

Wow, a very detailed analysis!

So, the question is whether we want to support old undocumented language-specific relations.

I would say not.

Do I understand correctly that the most sensible substitution would be to obl:mod? Currently I made it to obl.

[matyaskopp]

The most sensible substitution is advmod:lmod if ADV, obl:lmod otherwise. (http://hdl.handle.net/11346/PMLTQ-ORIW)

relation	pos	occurences
advmod:lmod	ADV	560
case	ADP	3
obl	NOUN	1
obl:lmod	NOUN	27
obl:lmod	ADP	15
obl:lmod	VERB	4
obl:lmod	ADJ	1
obl:lmod	X	1

[TomazErjavec]

The most sensible substitution is advmod:lmod if ADV, obl:lmod otherwise

Hm, maybe most correct, not sure about sensible, because the code now does not have access to the PoS of the word: https://github.com/clarin-eric/ParlaMint/blob/d02bd049213da4a3d1e50bea07df01215883fc0b/Scripts/parlamint2release.xsl#L562-L573.

Trying to implement PoS-dependnet dependency would be difficult. I would just set it to advmod:lmod, as this seems to mean only about 10% of errors. Which is about par on the error rate parsers make anyway...

[matyaskopp]

Trying to implement PoS-dependnet dependency would be difficult. I would just set it to advmod:lmod, as this seems to mean only about 10% of errors. Which is about par on the error rate parsers make anyway...

ok, but it will probably produce an L2 validation error - I think advmod should be related to ADV

[TomazErjavec]

ok, but it will probably produce an L2 validation error - I think advmod should be related to ADV

Ah. But given that we are patching things, might as well have some errrors...

[TomazErjavec]

Surprisingly, no CoNLL-I errors were produced. So, closing.