Mandarin Part-of-Speech Tagging and Discriminative Reranking

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 1093–1102, Prague, June 2007. c⃝2007 Association for Computational Linguistics Mandarin Part-of-Speech Tagging and Discriminative Reranking Zhongqiang Huang1 1Purdue University West Lafayette, IN 47907 zqhuang@purdue.edu Mary P. Harper1,2 2University of Maryland College Park, MD 20742 mharper@casl.umd.edu Wen Wang SRI International Menlo Park, CA 94025 wwang@speech.sri.com Abstract We present in this paper methods to improve HMM-based part-of-speech (POS) tagging of Mandarin. We model the emission prob- ability of an unknown word using all the characters in the word, and enrich the stan- dard left-to-right trigram estimation of word emission probabilities with a right-to-left prediction of the word by making use of the current and next tags. In addition, we utilize the RankBoost-based reranking algorithm to rerank the N-best outputs of the HMM- based tagger using various n-gram, mor- phological, and dependency features. Two methods are proposed to improve the gen- eralization performance of the reranking al- gorithm. Our reranking model achieves an accuracy of 94.68% using n-gram and mor- phological features on the Penn Chinese Treebank 5.2, and is able to further improve the accuracy to 95.11% with the addition of dependency features. 1 Introduction Part-of-speech (POS) tagging is potentially help- ful for many advanced natural language processing tasks, for example, named entity recognition, pars- ing, and sentence boundary detection. Much re- search has been done to improve tagging perfor- mance for a variety of languages. The state-of-the- art systems have achieved an accuracy of 97% for English on the Wall Street Journal (WSJ) corpus (which contains 4.5M words) using various mod- els (Brants, 2000; Ratnaparkhi, 1996; Thede and Harper, 1999). Lower accuracies have been reported in the literature for Mandarin POS tagging (Tseng et al., 2005; Xue et al., 2002). This is, in part, due to the relatively small size and the different annotation guidelines (e.g., granularity of the tag set) for the an- notated corpus of Mandarin. Xue at el. (2002) and Tseng at el. (2005) reported accuracies of 93% and 93.74% on CTB-I (Xue et al., 2002) (100K words) and CTB 5.0 (500K words), respectively, each us- ing a Maximum Entropy approach. The character- istics of Mandarin make it harder to tag than En- glish. Chinese words tend to have greater POS tag ambiguity than English. Tseng at el. (2005) reported that 29.9% of the words in CTB have more than one POS assignment compared to 19.8% of the English words in WSJ. Moreover, the morphological prop- erties of Chinese words complicate the prediction of POS type for unknown words. These challenges for Mandarin POS tagging suggest the need to develop more sophisticated methods. In this paper, we investigate the use of a discriminative reranking approach to in- crease Mandarin tagging accuracy. Reranking ap- proaches (Charniak and Johnson, 2005; Chen et al., 2002; Collins and Koo, 2005; Ji et al., 2006; Roark et al., 2006) have been successfully applied to many NLP applications, including parsing, named entity recognition, sentence boundary …

Content truncated to 500 words