Do pipeline components affect training accuracy?

[nick-antisoc]

I would like some info on how different components depend on each other during training.

Let's say for example I want to train a text categorizer textcat.
If I add to the pipeline a ner component, will this addition improve the final accuracy?

Does the trainable component use the ones before it during training in a way that would result to better accuracy?
Do only vectors/pretraining affect accuracy or the actual pipeline plays also an important role here?

Thanks,
Nick

[polm]

This is actually something we're working on making clearer at the moment, but in general pipeline components don't affect each other's accuracy.

1. Components generally listen to a tok2vec or transformer layer further up the pipeline, which is the main determiner of accuracy besides the model itself.
2. Components are potentially influenced by changes to the doc object if they modify the same fields; this comes up in NER, where it's common to have a rule-based system and a statistical one and their order matters.
3. Components can have internal models that aren't in the pipeline general. In the pretrained pipelines NER uses a separate tok2vec from the pipeline one.

In your case of textcat and NER: no, the NER annotations are not going to affect textcat.

Do only vectors/pretraining affect accuracy or the actual pipeline plays also an important role here?

For textcat, yes, basically only the input representations would make a difference.