Question about POS performance

[PengLU1101]

Hi,

I am trying to do some seqs tagging tasks.

For POS, did you ever run your model on Universal dependency POS-en dataset?
I want to get it as my baseline.

I used your code to test on that dataset, yet the results is 95.60%, which is lower than reuslt (95.80) of vanilla LSTM-CNN-CRF model.

I think contextual embs are supposed to improve the accuracy.

[LiyuanLucasLiu]

Thanks so much for asking. We haven't tried that dataset, it would be great if you can share the dataset in some way.

The co-training version of LM-LSTM-CRF has some issues and we're recently working on this part. We release a new repo w. char contextual embs (w.o. co-training). Maybe you can first try the following repo; after fixing these issues, we will let you know.

https://github.com/LiyuanLucasLiu/Vanilla_NER

[PengLU1101]

Thanks so much for asking. We haven't tried that dataset, it would be great if you can share the dataset in some way.

The co-training version of LM-LSTM-CRF has some issues and we're recently working on this part. We release a new repo w. char contextual embs (w.o. co-training). Maybe you can first try the following repo; after fixing these issues, we will let you know.

https://github.com/LiyuanLucasLiu/Vanilla_NER

Thanks for your reply, I will test data using the new repo.

For the UD dataset, http://universaldependencies.org/
And http://www.petrovi.de/data/lrec16.pdf