Chinese Word Segmentation for Sub-character Representation

Nowadays, bidirectional long short-term memory neural network(Bi-LSTM) becomes the main structure for Chinese word segmentation tasks, which can obtain text information with time series. As a sequence model, the training speed of Bi-STM is very slow, while dilated convolution neural networks(DCNN) have a natural advantage in it which is designed to obtain information with a long length. In this paper, the sub-character information is concatenated with the ordinary features to enrich the input. Multiple contrast experiments are designed to verify the effect of applying DCNN and adding Conditional Random Fields (CRF). Experiments on the four datasets in SIGHAN2005 show that DCNN structure can improve the word segmentation effect in terms of Fl value and efficiency. The main advantage of the DCNN is that the speed is greatly faster than Bi-LSTM.