search

hikopensource / DAVAR-Lab-OCR

OCR toolbox from Davar-Lab
Apache License 2.0
737 stars 157 forks source link

Questions about TRIE #81

Open xingjianz opened 2 years ago

xingjianz commented 2 years ago

Hi, I am very interested in the TRIE model. It was a very interesting design and a very good paper. I do have a couple questions regarding some design choices:

  1. Have you guys tried using transformers instead of LSTMs for text recognition and IE? If so, was the results not as impressive? If not, was it due to efficiency concerns?
  2. Is there any particular reason why you chose character level convolution as the encoding method? Thanks!
volcano1995 commented 2 years ago

Hi, I am very interested in the TRIE model. It was a very interesting design and a very good paper. I do have a couple questions regarding some design choices:

  1. Have you guys tried using transformers instead of LSTMs for text recognition and IE? If so, was the results not as impressive? If not, was it due to efficiency concerns?
  2. Is there any particular reason why you chose character level convolution as the encoding method? Thanks!

Hi, thanks for your attention.

  1. We not try to use transformer instead of LSTM, transformer is hot and effective yet difficult for training. For IE task, some dataset such as SROIE, wildreceipt is small and may not converge well. However, it is a good concern in future.
  2. Because some dataset such as EPHOIE need to classify each character.