From what I understand:
- a projective dependency parsers can only output projective tree
- Wikipedia says: "a large majority of sentences in most languages are projective, i.e. they do not contain discontinuities. Only about 15-25% of actual sentences contain a discontinuity"
Therefore, assuming that in English about 15-25% of actual sentences contain a discontinuity, the maximum accuracy (i.e. percentage of correct parse trees for entire sentences) that can achieve a projective dependency parser in English would be 75-85% (since it would fail for all non-projective sentences).
Is this reasoning correct?