Closed lexmosolov closed 4 years ago
See https://github.com/natasha/naeval , they added stanza a few days ago. (Russian comments on the comparison: https://github.com/natasha/naeval/issues/1 ) TLDR: Stanza has good and fast POS and DEP, but (relatively) bad and slow NER.
And what about Stanza models in Spacy pipeline? https://github.com/explosion/spacy-stanza
It's just a wrapper, as usual. You can use it to get predictions, but not train your spacy models with it.
Quality of syntax for https://github.com/buriy/spacy-ru/releases/tag/v2.3_pre1 :
data/grameval/news.json:
UAS 93.29
LAS 86.86
data/grameval/wiki.json:
UAS 83.74
LAS 72.16
data/grameval/fiction.json:
UAS 95.67
LAS 91.12
data/grameval/social.json:
UAS 79.62
LAS 70.03
data/grameval/poetry.json:
UAS 72.04
LAS 60.52
Benchmarked on https://github.com/natasha/naeval . If you would compare this to Stanza, this is very close (sometimes better, sometimes worse), but Spacy is ~3x faster.
Which library is better and when?