Unsupervised Joint Monolingual Character Alignment and Word Segmentation

We propose a novel Bayesian model for fully unsupervised word segmentation based on monolingual character alignment Adapted bilingual word alignment models and a Bayesian language model are combined through product of experts to estimate the joint posterior distribution of a monolingual character alignment and the corresponding segmentation. Our approach enhances the performance of conventional hierarchical Pitman-Yor language models with richer character-level features. In the conducted experiments, our model achieves an 88.6% word token f-score on the standard Brent version of the Bernstein-Ratner corpora. Moreover, on standard Chinese segmentation datasets, our method outperforms a baseline model by 1.9-2.9 f-score points.