Tuning models

[maciejbiesek]

Is there any option to tune the models (NER, POS) you provided on own corpora?

[ryszardtuora]

There is an option to train existing models further on data using spaCy cli train command. Just provide the name, or link to the model as the argument of the --base-model parameter. You will need to convert your data to JSON format using convert command.

This should work, with the exception of POS tagger for morfeusz-based version, which is not a spaCy component.

[maciejbiesek]

So, to sum up, we can tune eg. the NER model and POS tagger in the simplest form, but we cannot bias models that are morfeusz-based to our specific data?

[ryszardtuora]

You can tune NER in the morfeusz version, but you cannot do so for its POS tagger.

If it is the morfeusz tokenization that you're after, I suppose you could retrain the basic tagger, and then use it as a component in the pipeline.

Adding the ability to retrain the morfeusz-version tagger, would require more work, but we will consider this.

[maciejbiesek]

Ok, I see, thank you :)