Word Lattice Reranking for Chinese Word Segmentation and Part-of-Speech Tagging

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 385–392 Manchester, August 2008 Word Lattice Reranking for Chinese Word Segmentation and Part-of-Speech Tagging Wenbin Jiang † ‡ Haitao Mi † ‡ Qun Liu † †Key Lab. of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China ‡Graduate University of Chinese Academy of Sciences Beijing, 100049, China {jiangwenbin,htmi,liuqun}@ict.ac.cn Abstract In this paper, we describe a new rerank- ing strategy named word lattice reranking, for the task of joint Chinese word segmen- tation and part-of-speech (POS) tagging. As a derivation of the forest reranking for parsing (Huang, 2008), this strategy reranks on the pruned word lattice, which potentially contains much more candidates while using less storage, compared with the traditional n-best list reranking. With a perceptron classifier trained with local fea- tures as the baseline, word lattice rerank- ing performs reranking with non-local fea- tures that can’t be easily incorporated into the perceptron baseline. Experimental re- sults show that, this strategy achieves im- provement on both segmentation and POS tagging, above the perceptron baseline and the n-best list reranking. 1 Introduction Recent work for Chinese word segmentation and POS tagging pays much attention to discriminative methods, such as Maximum Entropy Model (ME) (Ratnaparkhi and Adwait, 1996), Conditional Ran- dom Fields (CRFs) (Lafferty et al., 2001), percep- tron training algorithm (Collins, 2002), etc. Com- pared to generative ones such as Hidden Markov Model (HMM) (Rabiner, 1989; Fine et al., 1998), discriminative models have the advantage of flexi- bility in representing features, and usually obtains almost perfect accuracy in two tasks. Originated by Xue and Shen (2003), the typ- ical approach of discriminative models conducts c⃝ 2008. Licensed to the Coling 2008 Organizing Com- mittee for publication in Coling 2008 and for re-publishing in any form or medium. segmentation in a classification style, by assign- ing each character a positional tag indicating its relative position in the word. If we extend these positional tags to include POS information, seg- mentation and POS tagging can be performed by a single pass under a unify classification framework (Ng and Low, 2004). In the rest of the paper, we call this operation mode Joint S&T. Experiments of Ng and Low (2004) shown that, compared with performing segmentation and POS tagging one at a time, Joint S&T can achieve higher accuracy not only on segmentation but also on POS tagging. Besides the usual local features such as the character-based ones (Xue and Shen, 2003; Ng and Low, 2004), many non-local features related to POSs or words can also be employed to improve performance. However, as such features are gener- ated dynamically during the decoding procedure, incorporating these features directly into the clas- sifier results in problems. First, the classifier’s fea- ture space will grow much rapidly, which is apt to overfit on training corpus. Second, the variance of non-local features caused by the model evolution during the training procedure will hurt the param- eter tuning. Last but …

Content truncated to 500 words