Discriminative Boosting from Dictionary and Raw Text - A Novel Approach to Build A Chinese Word Segmenter

Chinese word segmentation (CWS) is a basic and important task for Chinese information processing. Standard approaches to CWS treat it as a sequence labelling task. Without manually annotated corpora, these approaches are ineffective. When a dictionary is available, dictionary maximum matching (DMM) is a good alternative. However, its performance is far from perfect due to the poor ability on out-of-vocabulary (OOV) words recognition. In this paper, we propose a novel approach that integrates the advantages of discriminative training and DMM, to build a high quality word segmenter with only a dictionary and a raw text. Experiments in CWS on different domains show that, compared with DMM, our approach brings significant improvements in both the news domain and the Chinese medicine patent domain, with error reductions of 21.50% and 13.66%, respectively. Furthermore, our approach achieves recall rate increments of OOV words by 42.54% and 23.72%, respectively in both domains.