Representations for category disambiguation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 201–208 Manchester, August 2008 Representations for category disambiguation Markus Dickinson Indiana University Bloomington, IN 47405 md7@indiana.edu Abstract As it serves as a basis for POS tagging, cat- egory induction, and human category ac- quisition, we investigate the information needed to disambiguate a word in a lo- cal context, when using corpus categories. Specifically, we increase the recall of an error detection method by abstracting the word to be disambiguated to a represen- tation containing information about some of its inherent properties, namely the set of categories it can potentially have. This work thus provides insights into the rela- tion of corpus categories to categories de- rived from local contexts. 1 Introduction and Motivation Category induction techniques generally rely on local contexts, i.e., surrounding words, to cluster word types together (e.g., Clark, 2003; Sch¨utze, 1995), using information of a kind also found in human category acquisition tasks (e.g., Mintz, 2002, 2003). Such information is also at the core of standard part-of-speech (POS) tagging, or disambiguation, methods (see, e.g., Manning and Sch¨utze, 1999, ch. 10), with the contexts generally abstracted to POS tags. The contextual informa- tion is similar in both tasks because induction is founded in part upon the notion that local contexts are useful for disambiguation: one morphosyntac- tically clusters words which should have the same category in the same contexts. But which contexts count as being the “same”? And to what extent do categories based on context distributions resemble c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. corpus annotation categories? Since disambigua- tion is in some sense more primary, to begin to answer these questions we investigate which rep- resentations are effective for category disambigua- tion. Disambiguating a word’s category in context has of course been explored in other situations, especially POS tagging. Rarely, however, has it been shown as to which information is the most accurate at disambiguation and which information is absolutely necessary, without mixing these is- sues with other tagging issues, such as smoothing and unknown word tagging. We need techniques which isolate disambiguation, placing less empha- sis on generalizing contexts to new data. To de- termine the essential information needed for accu- rate disambiguation, we start with a precise model and generalize it. Changing the model in small ways and evaluating the resulting precision will in- dicate how particular aspects of the representation are contributing to successful disambiguation. The central question of this paper is: which representation (of a word and its context) indi- cates that two situations should be categorized the same? In this context, POS annotation error detec- tion provides an ideal setting to explore represen- tations for disambiguation. Error detection relies on the assumption that words should be annotated consistently—in other words, contexts are grouped which accurately identify the category of a word as being consistent—and it does this with an empha- sis on high precision. In essence, error detection already …

Content truncated to 500 words