Worse results on NER tags than LayoultLMv3

#7
by hal9003 - opened

For some reason the model performs worse than LayoultLMv3 is there any particular reason?

  • did you use segment position embeddings?
  • I guess that's to be expected given that the encoder of UDOP was not specifically pre-trained for discriminative tasks

Yes. I am following your notebook https://github.com/NielsRogge/Transformers-Tutorials/blob/master/UDOP/Fine_tune_UDOPEncoderModel_on_FUNSD_(HuggingFace_Trainer).ipynb
And from what I am seeing the results there are also lower on UDOP compared to LayoultLMv3

Interesting! I guess UDOP will be mainly useful for generative tasks

Sign up or log in to comment