Ukrainian language support in Flair

[alanakbik]

This issue tracks the progress of adding support for the Ukrainian language from lang-uk to Flair. We would like to add:

• Ukrainian Flair embeddings trained by @dchaplinsky and available here: forward and backward. Should be made loadable with embeddings = FlairEmbeddings('uk-forward')and embeddings = FlairEmbeddings('uk-backward')
• Ukrainian NER by @dchaplinsky, available here. Should be made loadable with tagger = SequenceTagger.load('ner-ukrainian')
• Ukrainian part-of-speech tagger by @dchaplinsky, available here. Should be made loadable with tagger = SequenceTagger.load('pos-ukrainian')
• Ukrainian NER dataset described here. Loadable as corpus = NER_UKRAINIAN(). Should be integrated only once version 2.0 is complete.
• Ukrainian Universal Dependency Treebank, loadable as corpus = UD_UKRAINIAN().

[dchaplinsky]

This is the code for the NER corpus I've used:
https://github.com/lang-uk/flair-ner/blob/main/train_base.py#L32

and the code for the POS corpus:
https://github.com/lang-uk/flair-pos/blob/main/train_grid.py#L21

I'll take a look if I have fixed split for ner hosted somewhere else

[stefan-it]

Really cool idea!

I had to do a lot of manual preprocessing steps to get NER working when evaluating the ELECTRA model:

https://github.com/stefan-it/ukrainian-electra/blob/main/download_prepare_data_ner.sh

[dchaplinsky]

Oh, @stefan-it thanks for reminding me. Totally forgot about fixed split.

On a separate topic. Would you like to try to train electra on a better quality ukrainian texts?

[stefan-it]

Hey @dchaplinsky , I currently have access to TPUs, so if you have texts available I would love to pretrain another model 🤗

[dchaplinsky]

Yes I do! Could you contact me at chaplinsky[dot]dmitry on gmail?

[dchaplinsky]

Hi @alanakbik and @stefan-it

I've just uploaded two bigger models for the Ukrainian language:
https://huggingface.co/lang-uk/flair-uk-forward-large
https://huggingface.co/lang-uk/flair-uk-backward-large

Those has hidden_size=2048 (in contrast to the 1024 of the original ones) and trained on my data + data from Stefan (54gb in total).

I've also trained a downstream NER model on them, and got a nice 1.5% improvement over the previous one, will publish it shortly.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.