SpaCy with German: Passive, relative clause are not being determined (token.dep_ == sbp, rc)

[dkharlanau]

How to reproduce the behaviour

import spacy

nlp = spacy.load("de_core_news_md"  ) 
for label in nlp.get_pipe("parser").labels:
    print(label, " -- ", spacy.explain(label)) 

it shows the available labels for dependency:
............

• sb -- subject
• sbp -- passivized subject (PP)**
• svp -- separable verb prefix
• rc -- relative clause

when I call the nlp function with the text:

doc = nlp("Der Brief wird geschrieben.")

for token in doc:
    print(token.text, 'with dependency:', token.dep_ )
 
doc = nlp("Der Brief ist geschrieben worden")

where Der Brief ist geschrieben worden, Der Brief wird geschrieben._ is in Passiv .

However, it shows that the object's dependency is sb -- subject and not sbp -- passivized subject (PP)

Der with dependency: nk
Brief with dependency: **sb**  
wird with dependency: ROOT
geschrieben with dependency: oc
. with dependency: punct

Der  / with dependency: nk
Brief  / with dependency: **sb**
ist  / with dependency: ROOT
geschrieben  / with dependency: oc
worden  / with dependency: oc

the same situation for relative clause

doc = nlp("Die Kinder, die dort spielen, wohnen in der Schweiz")

for token in doc:
    print(token.text, 'with dependency:', token.dep_ )

Die with dependency: nk
Kinder with dependency: sb
, with dependency: punct
**die with dependency: sb (??? ) -? rc**
dort with dependency: mo
spielen with dependency: rc
, with dependency: punct
wohnen with dependency: ROOT
in with dependency: mo
der with dependency: nk
Schweiz with dependency: nk

Version is 3.3.0
NAME SPACY VERSION
de_core_news_lg >=3.3.0.dev0,<3.4.0 3.3.0 ✔
de_core_news_md >=3.3.0.dev0,<3.4.0 3.3.0 ✔
de_core_news_sm >=3.3.0.dev0,<3.4.0 3.3.0 ✔
de_dep_news_trf >=3.3.0.dev0,<3.4.0 3.3.0 ✔
en_core_web_sm >=3.3.0.dev0,<3.4.0 3.3.0 ✔

Can you check what is the problem here? The result is the same independent on the model ( lg,md, trf ).

[adrianeboyd]

I think you're misunderstanding how these dependency labels are used.

rc is used to attach the head of the relative clause to the noun it's modifying ("Kinder" -> "spielen"):

sbp is used for the "von X" phrase (logical subject / agent) in a passive sentence, not the grammatical subject of the passive sentence ("geschrieben" -> "von"):

The grammatical subject ("Brief") still uses "sb" as in an active sentence.

Here's the full TIGER annotation manual (in German): https://www.ims.uni-stuttgart.de/documents/ressourcen/korpora/tiger-corpus/annotation/tiger_scheme-syntax.pdf

Obviously you'll run into some level of statistical errors in the models (see #3052), and especially for sentences that look less like the training data (newspaper text), but rc and sbp in these particular examples look correct to me.

[dkharlanau]

Thanks a lot for your quick response. With your comment, I agree now. Thank you for the clarification and the link to Tiger annotation.