Companion data necessary for training? about perin HOT 5 CLOSED

David knows better, but my feeling is that we use the lemmas from the companion data. I.e., when constructing the rules for labels, we allow copying/modifying a corresponding lemma instead of a token (or other sources).

So either you need to lemmatize the data (you could use the UDPipe service, for example, we have the new Bert version trained on UD 2.6 running on https://lindat.mff.cuni.cz/services/udpipe/), or you could disable the usage of the lemma rules (and perform the lemmatization during the syntactic parsing).

from perin.

Yeah, the only problem is in the lemmatized tokens, which are used to create more efficient set of relative label rules -- so specifically, the absence of the companion data shouldn't impact UCCA parsing (but it will most likely negatively influence the accuracy of label prediction for the other frameworks).

I've quickly hacked a workaround to preprocess the data without a companion file into the branch no_lemmas.

from perin.

As for generating the companion data (i.e. lemmas), you can use the code from UDPipeWrapper.

from perin.

This makes a lot of sense. Thank you both for the quick solution!
Besides, it's good to know the new UDPipe is already usable, even if not yet offline.
I'll be happy to close the issue unless you want to keep it, e.g. for adding documentation about this option.

from perin.

Merged into the main branch [#9], closing.

from perin.