We use Python3.6+tensorflow1.12.0 to coding the bi-LSTM+CRF network structure and realizing the sequence tagging for chinese characters.
(1)Raw data preprocessing. Calling vocab_build() function to convert the data of .txt format to .pkl format. And initilazing the word vectors.(We set the dimension of vectors is 300)
(2)Designing the network structure and hyper-parameters, calling main.py train and test model.
- batch_size=64;
- epoch=25;
- learning_rate=0.001;
- dropout=0.5;
- gradient_clipping=5.0;
- LSTM_num(forward)=300;
- LSTM_num(backward)=300；
- optimizer=Adam;
- ...
After 25 epochs, the precision, recall, F1 values of test set as shoun below.
Fig. precision, recall, F1 values
- precision=0.951266;
- recall=0.908613;
- f1=0.929270. (After 25 iterations)
If you want to know more about the LSTM+CRF model, please to read sequence labelling.md or click here.