POS Tagging with a More Informative Tagset

Proceedings of the Australasian Language Technology Workshop 2005, pages 40–48, Sydney, Australia, December 2005. POS Tagging with a More Informative Tagset Andrew MacKinlay† and Timothy Baldwin‡ †‡ Dept. of Computer Science and Software Engineering University of Melbourne Victoria 3010 Australia ‡ NICTA Victoria Research Lab University of Melbourne Victoria 3010 Australia {amack,tim}@cs.mu.oz.au Abstract We investigate the impact of introducing finer dis- tinctions into the tagset on the accuracy of part- of-speech tagging. This is a tangential approach to most recent research in the field, which has focussed on applying different algorithms using a very simi- lar set of features. We outline the basic approach to tagset refinement and describe preliminary find- ings. 1 Introduction Most recent research on corpus-based part-of- speech (POS) tagging has tended to focus on applying new algorithms to an existing task (Brill, 1995; Ratnaparkhi, 1996; Gim´enez and M`arquez, 2004), or improving the efficiency of an existing algorithm (Ngai and Florian, 2001). While there has been some successful experi- mentation with modifying the feature sets of particular taggers (Toutanova and Manning, 2000), the various state-of-the-art taggers for the most part use a very similar set of features in determining the tag for a particular token: some subset of the two preceding and two fol- lowing tokens, and their tags. The different al- gorithms have tended to plateau to a similar “glass ceiling” in accuracy (96.9 ± 0.3% over all tokens for the taggers in this paper) by using these features. POS tagging is essentially an optimisation process over firstly the tag sequence and sec- ondly the tag–word assignments for a given in- put. The relative difficulty of this task hinges on the internals of the POS tagset, and the tag- ging performance over a given dataset can vary greatly depending on the tagset used. In this paper, we seek to enhance tagging performance by adding a third dimension to the optimisation process: the tagset. We explore the possibility of transforming the tagset via reversible (loss- less) mappings, to produce a dataset which is more amenable to automatic tagging and thus results in higher performance than the origi- nal tagset. We follow the majority of recent mainstream research on English POS tagging in adopting as a baseline the tagset used in the Penn Treebank (Marcus et al., 1993). We evaluate two different approaches to iden- tifying patterns of syntactic regularity with the existing POSs. Our primary approach is lin- guistic insight: we investigate a range of linguis- tically motivated subdivisions which are either designed to assist in a specific problematic in- stance of disambiguation, or are linguistically sensible enough to be applied for independent reasons. Additionally we compare this with a data-driven approach, where we attempt to identify intra-POS groupings by running a clus- tering algorithm over the words within a partic- ular class using features derived from syntactic context. We report the most promising results achieved in both cases. Section 2 outlines some diverse algorithms which have been applied to POS tagging; Sec- tion 3 gives some …

Content truncated to 500 words