Inaccurate tagging of csubjpass

[fatihbozdag]

Greetings all,

I am working on extracting subordinate clauses with Spacy. However, dependency parsing provides incorrect results for csubjpass if there is another word prior to passive verb.

For instance:

doc_1 = nlp('what she said was well received')
for t in doc_1:
    print(t.text, t.dep_, t.head.text)

what dobj said
she nsubj said
said csubj was
was ROOT was
well advmod received
received acomp was

However,

doc_2 = nlp('what she said was received')
for t in doc_1:
    print(t.text, t.dep_, t.head.text)

what dobj said
she nsubj said
said csubjpass received
was auxpass received
received ROOT received

In first case, said should have been tagged as csubjpass, as exemplified here:

How should I proceed, then? Any suggestions?

Spacy ver: 3.4
Spacy Modal: en_core_web_trf

[polm]

If this issue is systematic like you describe, I would recommend writing a small custom component to fix the annotations. You could also probably use the AttributeRuler.

In addition to #3052 about errors in general, I suspect the model has trouble with this because this kind of usage isn't very common in the training data (or in many kinds of writing, for that matter).

[fatihbozdag]

Well, strange thing is that Stanza and UD Models also suffer the same issue. I'd better start working on AttributeRuler.