which/that... as ADJ or DET. Wouldn't PRON fit better?

[c0ntradicti0n]

How to reproduce the behaviour

If find it strange and annoying, that "which" and "that" are no pronouns and are not recognized by for example Tom Wolfs neuralcoref, for I assume this reason. Wouldn't that be nice, if you can say, to what word this relative pronoun is related to?

https://explosion.ai/demos/displacy?text=There%20are%20theories%2C%20that%20would%20say%2C%20that%20classifying%20%20%22which%22%20and%20%22that%22%20as%20ADJECTIVES%20or%20DETERMINERS%2C%20is%20a%20behavior%2C%20which%20is%20not%20quite%20convincing.&model=en_core_web_lg&cpu=1&cph=1

Your Environment

https://universaldependencies.org/u/pos/all.html#al-u-pos/PRON

[honnibal]

My kneejerk reaction here was "well that's the annotation scheme", but looking at the UD link, maybe you're right and we should just update the tag map? I think it'll mean the neural coref models have to be updated though, so it's something we should be a bit careful with.

Do you know whether the coreference data has annotations for relative pronouns?

[c0ntradicti0n]

It seems like there are no annotations in this data. I looked here: http://conll.cemantix.org/2012/download/conll-2012-trial-data.tar.gz That't a public dataset that Thomas Wolf refers to.
There are cases like which as WDT:

bc/phoenix/00/phoenix_0002          1          0         In         IN (TOP(S(PP*          -          -          -  speaker#2          *          * (ARGM-DIS*          *          *          *          -
bc/phoenix/00/phoenix_0002          1          1       fact         NN     (NP*))          -          -          -  speaker#2          *          *         *)          *          *          *          -
bc/phoenix/00/phoenix_0002          1          2          ,          ,          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1          3         on         IN       (PP*          -          -          -  speaker#2          *          * (ARGM-TMP*          *          *          *          -
bc/phoenix/00/phoenix_0002          1          4     August        NNP    (NP(NP*          -          -          -  speaker#2     (DATE*          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1          5         20         CD         *)          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1          6          ,          ,          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1          7       1976         CD    (NP*)))          -          -          -  speaker#2         *)          *         *)          *          *          *          -
bc/phoenix/00/phoenix_0002          1          8          ,          ,          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1          9        the         DT       (NP*          -          -          -  speaker#2      (ORG*          *     (ARG0*          *          *          *          -
bc/phoenix/00/phoenix_0002          1         10      Hebei        NNP          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         11    Science        NNP          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         12  Committee        NNP         *)          -          -          -  speaker#2         *)          *         *)          *          *          *          -
bc/phoenix/00/phoenix_0002          1         13        had        VBD       (VP*       have         01          -  speaker#2          *       (V*)          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         14     issued        VBN       (VP*      issue         01          1  speaker#2          *          *       (V*)          *          *          *          -
bc/phoenix/00/phoenix_0002          1         15          a         DT    (NP(NP*          -          -          -  speaker#2          *          *     (ARG1*     (ARG0*          *          *        (56
bc/phoenix/00/phoenix_0002          1         16      brief         JJ          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         17     report         NN         *)          -          -          -  speaker#2          *          *          *         *)          *          *          -
bc/phoenix/00/phoenix_0002          1         18      which        WDT (SBAR(WHNP*)          -          -          -  speaker#2          *          *          *  (R-ARG0*)          *          *          -
bc/phoenix/00/phoenix_0002          1         19        for         IN     (S(PP*          -          -          -  speaker#2          *          *          * (ARGM-TMP*          *          *          -
bc/phoenix/00/phoenix_0002          1         20        the         DT       (NP*          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         21      first         JJ          *          -          -          -  speaker#2  (ORDINAL)          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         22       time         NN        *))          -          -          -  speaker#2          *          *          *         *)          *          *          -
bc/phoenix/00/phoenix_0002          1         23   revealed        VBD       (VP*     reveal         01          2  speaker#2          *          *          *       (V*)          *          *          -
bc/phoenix/00/phoenix_0002          1         24        the         DT    (NP(NP*          -          -          -  speaker#2          *          *          *     (ARG1*          *          *          -
bc/phoenix/00/phoenix_0002          1         25      facts        NNS         *)          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         26      about         IN       (PP*          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         27   Qinglong         NN    (S(NP*)          -          -          -  speaker#2      (GPE)          *          *          *          *    (ARG0*)       (13)
bc/phoenix/00/phoenix_0002          1         28     having        VBG       (VP*       have         01         12  speaker#2          *          *          *          *       (V*)          *          -
bc/phoenix/00/phoenix_0002          1         29 successfully         RB    (ADVP*)          -          -          -  speaker#2          *          *          *          *          * (ARGM-MNR*)          -
bc/phoenix/00/phoenix_0002          1         30   prepared        VBN       (VP*    prepare         02          2  speaker#2          *          *          *          *          *       (V*)          -
bc/phoenix/00/phoenix_0002          1         31    against         IN       (PP*          -          -          -  speaker#2          *          *          *          *          *     (ARG2*          -
bc/phoenix/00/phoenix_0002          1         32        the         DT       (NP*          -          -          -  speaker#2    (EVENT*          *          *          *          *          *        (73
bc/phoenix/00/phoenix_0002          1         33      Great        NNP          *          -          -          -  speaker#2          *          *          *          *          *          *          -
bc/phoenix/00/phoenix_0002          1         34   Tangshan        NNP          *          -          -          -  speaker#2          *          *          *          *          *          *       (91)
bc/phoenix/00/phoenix_0002          1         35 Earthquake        NNP *)))))))))))))          -          -          -  speaker#2         *)          *         *)         *)          *         *)    73)|56)
bc/phoenix/00/phoenix_0002          1         36          .          .        *))          -          -          -  speaker#2          *          *          *          *          *          *          -

I think, there is another reason for the missing annotations. It is clear from the syntax by the dependence of the verb in the subordinate clause, to what that leading pronoun refers to. https://explosion.ai/demos/displacy?text=I%20love%20you%2C%20whose%20brother%20is%20married%20to%20my%20sister.&model=en_core_web_sm&cpu=1&cph=1
They skipped that presumably, because they just wanted to annotate the unclear things. So the models of Thomas Wolf could be updated by a simple algorithmic thing, but evaluation should refer to these 'unclear' relations without that. But if the tags are like 'ADJ' it's a little bit more difficult, but for sure, that would be possible without a special tag.

[thomwolf]

Hi guys, I really want to update Neuralcoref soon to fix the installation issues (it's in a bad shape right now). It would be a good thing to check this tag question at the same time. Currently plan is to work on that early March.

[honnibal]

Argh. This was actually in a patch that got orphaned from 1.x -- it never got applied to v2.0...And then I didn't read the issue correctly that pointed this out, and it never came up as a PR. Sorry!

Anyway. Fixed now. And we're winning the war on the issue tracker, so the same thing shouldn't happen in future :)

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.