Variety in lemmatizer output

[spatiebalk]

I am training a NER model that also needs to have a lemmatizer. This lemmatizer I initialized with the source model en_core_web_sm as that is fine and does not need to be trained in my use case. The NER part works well, however, I have some questions about some of the behaviour of the lemmatizer.

1. I initialized the lemmatizer with the one from en_core_web_sm, and froze it during training, however I see that sometimes my trained model assigns a different lemma to a token then the en_core_web_sm model does.
2. For my trained model (and not trained lemmatizer) I see a varying performance of the lemmatizer where the same token gets different lemma values in different sentences.

So I was wondering, can this behaviour be explained by the lemmatizer being context dependent or are these artifacts of some mistake during training?
For completeness, I have added the config.cfg file I have used.

[thomashacker]

Hello,
you also need to freeze the tok2vec component, since all other components sourced from en_core_web_sm are listening to it.
The lemmatizer depends on the tagger, so you can also verify whether the POS tags are also correct.