An Empirical Study Of Semi-Supervised Chinese Word Segmentation Using Co-Training

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1191–1200, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics An Empirical Study Of Semi-Supervised Chinese Word Segmentation Using Co-Training Fan Yang Nuance Communications, Inc. fan.yang@nuance.com Paul Vozila Nuance Communications, Inc. paul.vozila@nuance.com Abstract In this paper we report an empirical study on semi-supervised Chinese word segmenta- tion using co-training. We utilize two seg- menters: 1) a word-based segmenter lever- aging a word-level language model, and 2) a character-based segmenter using character- level features within a CRF-based sequence labeler. These two segmenters are initially trained with a small amount of segmented data, and then iteratively improve each other using the large amount of unlabelled data. Our experimental results show that co-training captures 20% and 31% of the performance improvement achieved by supervised training with an order of magnitude more data for the SIGHAN Bakeoff 2005 PKU and CU corpora respectively. 1 Introduction In the literature there exist two general models for supervised Chinese word segmentation, the word- based approach and the character-based approach. The word-based approach searches for all possible segmentations, usually created using a dictionary, for the optimal one that maximizes a certain util- ity. The character-based approach treats segmenta- tion as a character sequence labeling problem, indi- cating whether a character is located at the bound- ary of a word. Typically the word-based approach uses word level features, such as word n-grams and word length; while the character-based approach uses character level information, such as character n- grams. Both approaches have their own advantages and disadvantages, and there has been some research in combining the two approaches to improve the per- formance of supervised word segmentation. In this research we are trying to take advantage of the word-based and the character-based approaches in the semi-supervised setting for Chinese word seg- mentation, where there is only a limited amount of human-segmented data available, but there ex- ists a relatively large amount of in-domain unseg- mented data. The goal is to make use of the in- domain unsegmented data to improve the ultimate performance of word segmentation. According to Sun et al. (2009), “the two approaches [word-based and character-based approaches] are either based on a particular view of segmentation.” This naturally motivates the use of co-training, which utilizes two models trained on different views of the input la- beled data which then iteratively educate each other with the unlabelled data. At the end of the co- training iterations, the initially weak models achieve improved performance. Co-training has been suc- cessfully applied in many natural language process- ing tasks. In this paper we describe an empiri- cal study of applying co-training to semi-supervised Chinese word segmentation. Our experimental re- sults show that co-training captures 20% and 31% of the performance improvement achieved by super- vised training with an order of magnitude more data for the SIGHAN Bakeoff 2005 PKU and CU corpora respectively. In section 2 we review the two supervised ap- proaches and co-training algorithm in …

Content truncated to 500 words