Improving Chinese POS Tagging with Dependency Parsing

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1447–1451, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Improving Chinese POS Tagging with Dependency Parsing Zhenghua Li, Wanxiang Che, Ting Liu∗ Research Center for Social Computing and Information Retrieval MOE-Microsoft Key Laboratory of Natural Language Processing and Speech School of Computer Science and Technology Harbin Institute of Technology, China {lzh,car,tliu}@ir.hit.edu.cn Abstract Recent research usually models POS tag- ging as a sequential labeling problem, in which only local context features can be used. Due to the lack of morphological inflections, many tagging ambiguities in Chinese are difficult to handle unless con- sulting larger contexts. In this paper, we try to improve Chinese POS tagging by us- ing long-distance dependencies produced by a statistical dependency parser. Ex- perimental results show that, despite er- ror propagation, the syntactic features can significantly improve the tagging accuracy from 93.88% to 94.41% (p < 10−5). Detailed analysis shows that these fea- tures are helpful for ambiguous pairs like {NN,VV} and {DEC,DEG}.1 1 Introduction Part-of-speech (POS) tagging is a necessary and important step for many natural language tasks, for example, named entity recognition, parsing and sentence boundary detection. In the cur- rent literature, POS tagging is treated as a typ- ical sequence labeling problem, to which many models have been successfully applied, such as maximum-entropy (Ratnaparkhi, 1996), condi- tional random fields (CRF) (Lafferty et al., 2001) and perceptron (Collins, 2002). To facilitate fast decoding, these models make the Markovian inde- pendence assumption that the current tag depends on one or two previous tags. In addition, they can merely consider local context features, e.g. two ∗Correspondence author: tliu@ir.hit.edu.cn 1DEG and DEC are the two POS tags for the frequently used auxiliary word “�” (d¯e, of) in Chinese. The associative “�” is tagged as DEG, such as “I �/father � � �/eyes (eyes of the father)”; while the one in a relative clause is tagged as DEC, such as “�/he � �/made � ? �/progress (progress that he made)”. words in both sides of the focus word. This works quite well for English, because inflections are use- ful and strong clues for POS tags. However, due to the lack of morphological inflections, Chinese POS tagging has proven to be much more chal- lenging than English. With a typical sequential la- beling model such as Conditional Random Fields (CRF), the tagging accuracy is about 97% for En- glish, while less than 94% for Chinese (Huang et al., 2009). NN-VV ambiguities are one of the most noto- rious difficulties for Chinese POS tagging. Figure 1 gives two examples. We can see that the POS tagger can effortlessly assign the right tags to both “development” and “develop” in the English side. However, it is very difficult in the Chinese side since no word form inflection is available and the context features may be too sparse or uninforma- tive. However, the introduction of long-distance dependencies can largely reduce this difficulty. In the upper example of Figure 1, the …

Content truncated to 500 words