truncate a long sentence in backpropagation about sequence_tagging

[guillaumegenthial]

I assume you refer to this page (https://www.tensorflow.org/tutorials/recurrent#truncated-backpropagation).
For CRF, as the score is at the sequence level, that might be harder to perform. It would require passing through the dynamic programming stuff... (would require a lot of work) so I don't really know how i could help you. I see a few options:

1. either you let go of the CRF
2. you train the LSTM first and then retrain with the CRF layer without truncating
3. you train the LSTM and CRF separately and merge as in step 2. To train the CRF separately you could just store the outputs of the LSTM that you get from truncation and then feed them directly to the CRF layer as feed_dict...

hope that helps

[bioRichard]

@guillaumegenthial thank you I'll try your advice.

[guillaumegenthial]

if you come to a solution, please share it, I'm sure it will help others!