Improving Part-of-speech Tagging for Context-free Parsing

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1260–1268, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Improving Part-of-speech Tagging for Context-free Parsing Xiao Chen and Chunyu Kit Department of Chinese, Translation and Linguistics City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong SAR, P. R. China {cxiao2,ctckit}@cityu.edu.hk Abstract In this paper, we propose a factored pars- ing model consisting of a lexical and a constituent model. The discriminative lex- ical model allows the parser to utilize rich contextual features beyond those encoded in the context-free grammar (CFG) in use. Experiment results reveal that our parser achieves statistically significant improve- ment in both parsing and tagging accuracy on both English and Chinese. 1 Introduction Part-of-speech (POS) tagging, or tagging for short, is usually considered a front-end preparation for parsing. Folk wisdom holds that accurate tagging result is helpful to alleviate the syntactic ambigu- ity problem in parsing, motivating a huge amount of research on perfecting tagging techniques. In the past two decades, most POS tagging systems were based on a sequential classifica- tion approach, decomposing a sequence label- ing task into a series of classification subtasks. The state of the art of tagging was achieved by virtue of well-developed machine learning meth- ods, e.g. the Maximum Entropy model as in Rat- naparkhi (1996) and the Support Vector Machine as in Gimenez and Marquez (2003). All these techniques boosted the performance of POS tag- ging significantly. The error rate of the best En- glish POS tagger is less than 3% (Ratnaparkhi, 1996; Toutanova et al., 2003; Shen et al., 2007), very close to the inter-annotator discrepancy of the Penn Treebank (Marcus et al., 1993). For a long time, however, parsing seems to have been evolving in parallel to tagging with- out much interaction with each other. Generative parsers (Collins, 1997; Charniak, 2000; Charniak and Johnson, 2005; Petrov and Klein, 2007), ow- ing to their generative nature, all include a lexi- cal probability model in the form P(w|t). The in- formation used to predict the POS tag of a word mainly comes from the word itself and/or the an- cestors that derive this word in a tree. Local con- text, which is proved to be the most useful in- formation source for tagging (Ratnaparkhi, 1996; Toutanova and Manning, 2000), is not efficiently utilized by these parsers. Another noticable fact is that these high-performance parsers cannot do a better job of POS tagging than most of the state- of-the-art taggers (see Section 4 for a comparison). This is quite against our intuition that a parser hav- ing access to syntactic and contextual information from all over an input sentence should outperform a tagger that is limited to utilizing only local con- text and sequential dependency. This is an obser- vation in Charniak et al. (1996). But their explana- tion is that a parser is built to find phrase markers, not tags. Then an interesting question arises: if a parser cannot tag words better, how can …

Content truncated to 500 words