hankcs / ID-CNN-CWS

Source codes and corpora of paper "Iterated Dilated Convolutions for Chinese Word Segmentation"

ID-CNN-CWS

Source codes and corpora of paper "Iterated Dilated Convolutions for Chinese Word Segmentation" published in NNW journal.

It implements the following 4 models for CWS:

• Bi-LSTM
• Bi-LSTM-CRF
• ID-CNN
• ID-CNN-CRF

Dependencies

• Python >= 3.6
• TensorFlow >= 1.2

Both CPU and GPU are supported. GPU training is 10 times faster.

Preparation

Run following script to convert corpus to TensorFlow dataset.

$ ./scripts/make.sh

Train and Test

Quick Start

$ ./scripts/run.sh $dataset $model

• $dataset can be pku, msr, asSC or cityuSC.
• $model can be cnn or bilstm.

For example:

$ ./scripts/run.sh pku cnn

It will train a cnn model on pku dataset, then evaluate performance on test set.

CRF Layer

To enable CRF layer, simply append --viterbi to your command, e.g.

$ ./scripts/run.sh pku cnn --viterbi

Accuracy

Speed

Acknowledgments

• Corpora are from SIGHAN05, converted to Simplified Chinese via HanLP. Note that the SIGHAN datasets should only be used for research purposes.
• Model implementations adopted from https://github.com/iesl/dilated-cnn-ner by Emma Strubell.