BiLSTM-CNN-CRF v. 2.2.0

BiLSTM-CNN-CRF implementation using Python 3, Keras 2.2.0 and Tensorflow 1.8.0.

Note, it is not clear whether the code will run with different versions of Keras or Tensorflow or with different keras backends than Tensorflow.

Changes to previous version:

• Update of Keras & Tensorflow to latest version
• Bugfix for loading models that used character based word representations
• Models trained from the previous version that use CharEmbeddings: 'CNN' cannot be loaded with this version. This is due to the fact that Keras / Tensorflow no longer support zero masking for conv1d layers. Such models must be retrained with the new system.

BiLSTM-CNN-CRF v. 2.1.5

BiLSTM-CNN-CRF implementation using Python 3, Keras 2.1.5 and Tensorflow 1.7.0.

Note, it is not clear whether the code will run with different versions of Keras or Tensorflow or with different keras backends than Tensorflow.

BiLSTM-CNN-CRF v. 1.2.2

BiLSTM-CNN-CRF implementation using Keras 1.2.2, Theano 0.9.0, and Tensorflow 0.12.1.

Note: The implementation can only run with these versions of Keras/Theano/Tensorflow. It does not work with more recent versions.
It is recommended to use Keras with Theano as a backend.