Ambiguous dependency parsing for negation

[mbrine555]

When using the dependency parser, it seems like there can be a lot of ambiguity when trying to assign negation, unlike with something like Stanford's parser.

For example:

import spacy
nlp = spacy.load("en_core_web_sm")
sentence = u"The camera is not pretty, but it is functional."
doc = nlp(sentence)

for token in doc:
    print(token.dep_ + "(" + token.head.text + ", " + token.text + ")")

returns

det(camera, The)
nsubj(is, camera)
ROOT(is, is)
neg(is, not)
acomp(is, pretty)
punct(is, ,)
cc(is, but)
nsubj(is, it)
conj(is, is)
acomp(is, functional)
punct(is, .)

This output seems to imply that the negation could refer to either acomp, which is not the case. A Stanford output for the same sentence looks something like:

det(camera-2, The-1)
nsubj(pretty-5, camera-2)
cop(pretty-5, is-3)
neg(pretty-5, not-4)
root(ROOT-0, pretty-5)
cc(pretty-5, but-7)
nsubj(functional-10, it-8)
cop(functional-10, is-9)
conj(pretty-5, functional-10)

Is there any way to clear up this ambiguity currently?

Your Environment

• Python Version Used: 2.7.14
• spaCy Version Used: 2.0.11

[honnibal]

@mbrine555 Which Stanford dependency scheme is that?

[mbrine555]

@honnibal Sorry, I should've been more clear. It's the Universal Dependency scheme used in Stanford's CoreNLP 3.9.1.

[dxiao2003]

If you run the visualizer you can see that the "not" refers specifically to the first "is" token. Not exactly the same parse as Stanford but at least it's not ambiguous as to which "is" it's referring to.

[ines]

Merging this with #3052. We've now added a master thread for incorrect predictions and related reports – see the issue for more details.

[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.