A Local Generative Model for Chinese Word Segmentation

This paper presents a local generative model for Chinese word segmentation, which has faster learning process than discriminative models and can do unsupervised learning. It has the ability to make use of larger resources. In this model, four successive characters are used to determine whether a character interval should be a word boundary or riot. The Gibbs sampling algorithm, as well as three additional rules, is applied for the unsupervised learning. Besides words, the word candidates that are generated by our model can improve the performance of Chinese information retrieval. The experiments show that in supervised learning our method outperforms a language model based method. And the performance on one corpus is better than the best one reported in SIGH:AN bakeoff 05. In unsupervised learning, our method achieves the comparable performance compared to the state-of-the-art method.