Chinese word segmentation in tensorflow.

The architecture of this model is simple. There are three components of the model:

• Embedding: words embedding layer
• BiLSTM: a bidirectional LSTM layer
• CRF: a conditional random field layer

Segmentation is some kind of tagging. We can tag each token of the input sequence with Only a few tags:

• B: begin of a token
• M: middle of a token
• E: end of a token
• S: single character as a token
• O: Out of tags

We train the model to tag every input sequence, and then wo process the tagged result, so we get the final segmentation.

Assuming that we have a hparams file in deepseg/example_params.json:

python -m deepseg.runner \
    --params_file=deepseg/example_params.json \
    --mode=train

python -m deepseg.runner \
    --params_file=deepseg/example_params.json \
    --mode=eval

python -m deepseg.runner \
    --params_file=deepseg/example_params.json \
    --mode=predict

python -m deepseg.runner \
    --params_file=deepseg/example_params.json \
    --mode=train_and_eval

You may want to export the model to saved model format and serve it on tf serving, you can just run:

python -m deepseg.runner \
    --params_file=deepseg/example_params.json \
    --mode=export