The dependency parse result is different between displacy and local environment.

[meteorlin]

Hello!
I have installed spaCy (3.3.0) and used "en_core_web_??" (3.3.0) in my local environment.

But when I test the text of "Save Mary Jane" under the model setting of "en_core_web_sm" or "en_core_web_md", the dependency parser give me the wrong result which mistake "Save" as "PROPN".

import spacy
nlp = spacy.load("en_core_web_sm")  # or "en_core_web_md"
doc = nlp("Save Mary Jane")
spacy.displacy.serve(doc, style="dep")

For comparison, I test it in online displacy demo, but it give me the right answer under the model setting of "en_core_web_sm".

After these experiments, I try "en_core_web_lg" in local environment and it gave me the same answer as online displacy demo gave me.

So, is "en_core_web_lg" model used in the online displacy demo? Or just my settings are wrong or miss other important settings in my local environment?
Thanks in advance!

[polm]

As visible in your screenshot, the online demo is using 3.1.0, which is why it's different from your local 3.3.0 output.

[meteorlin]

Oh! I reinstall the spaCy and en_core_web_sm in 3.1.0 version and it works fine. But I am so curious that why the latest model will perform worse than the previous?

[polm]

As noted in #3052, the models are statistical and explaining individual predictions is difficult. In this case it's likely the model is "worse" in this particular case but better in others.