search

heshenghuan / LSTM-CRF

A (CNN+)RNN(LSTM/BiLSTM)+CRF model for sequence labelling.:smirk:
140 stars 58 forks source link

about Usage #1

Closed lin520chong closed 6 years ago

lin520chong commented 7 years ago

when can you write the Usage?

heshenghuan commented 7 years ago

@lin520chong sorry, I'm not sure... Because this project is unfinished. It will takes at most 2 weeks to be finished, I think... 😅

lin520chong commented 7 years ago

@heshenghuan waiting for you I'm also doing some work about LSTM for NER recently. Maybe we can share our ideas when you finish this project.

heshenghuan commented 7 years ago

@lin520chong sorry, I'm tortured by some Keras bugs recently, because I found that the same code will generate different result on CPU and GPU. That's really disappoint me, cause I cannot find any logical error on my code. So I decided to use tensorflow to implement a LSTM model. And it works well, I'm runing a test now.

And inspired by some LSTM Chinese word segmentation (check this: kcws), I no longer use single LSTM layer with 'softmax' activation, but added a CRF at the end of activation layer. I think this might be helpful to do structured predition.