A Cascaded Linear Model for Joint Chinese Word Segmentation and Part-of-Speech Tagging

Proceedings of ACL-08: HLT, pages 897–904, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics A Cascaded Linear Model for Joint Chinese Word Segmentation and Part-of-Speech Tagging Wenbin Jiang † Liang Huang ‡ Qun Liu † Yajuan L¨u † †Key Lab. of Intelligent Information Processing ‡Department of Computer & Information Science Institute of Computing Technology University of Pennsylvania Chinese Academy of Sciences Levine Hall, 3330 Walnut Street P.O. Box 2704, Beijing 100190, China Philadelphia, PA 19104, USA jiangwenbin@ict.ac.cn lhuang3@cis.upenn.edu Abstract We propose a cascaded linear model for joint Chinese word segmentation and part- of-speech tagging. With a character-based perceptron as the core, combined with real- valued features such as language models, the cascaded model is able to efficiently uti- lize knowledge sources that are inconvenient to incorporate into the perceptron directly. Experiments show that the cascaded model achieves improved accuracies on both seg- mentation only and joint segmentation and part-of-speech tagging. On the Penn Chinese Treebank 5.0, we obtain an error reduction of 18.5% on segmentation and 12% on joint seg- mentation and part-of-speech tagging over the perceptron-only baseline. 1 Introduction Word segmentation and part-of-speech (POS) tag- ging are important tasks in computer processing of Chinese and other Asian languages. Several mod- els were introduced for these problems, for example, the Hidden Markov Model (HMM) (Rabiner, 1989), Maximum Entropy Model (ME) (Ratnaparkhi and Adwait, 1996), and Conditional Random Fields (CRFs) (Lafferty et al., 2001). CRFs have the ad- vantage of flexibility in representing features com- pared to generative ones such as HMM, and usually behaves the best in the two tasks. Another widely used discriminative method is the perceptron algo- rithm (Collins, 2002), which achieves comparable performance to CRFs with much faster training, so we base this work on the perceptron. To segment and tag a character sequence, there are two strategies to choose: performing POS tag- ging following segmentation; or joint segmentation and POS tagging (Joint S&T). Since the typical ap- proach of discriminative models treats segmentation as a labelling problem by assigning each character a boundary tag (Xue and Shen, 2003), Joint S&T can be conducted in a labelling fashion by expand- ing boundary tags to include POS information (Ng and Low, 2004). Compared to performing segmen- tation and POS tagging one at a time, Joint S&T can achieve higher accuracy not only on segmentation but also on POS tagging (Ng and Low, 2004). Be- sides the usual character-based features, additional features dependent on POS’s or words can also be employed to improve the performance. However, as such features are generated dynamically during the decoding procedure, two limitation arise: on the one hand, the amount of parameters increases rapidly, which is apt to overfit on training corpus; on the other hand, exact inference by dynamic program- ming is intractable because the current predication relies on the results of prior predications. As a result, many theoretically useful features such as higher- order word or POS n-grams are difficult to be in- corporated in the model efficiently. To cope …

Content truncated to 500 words