Functional Elements and POS Categories

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1198–1206, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Functional Elements and POS Categories Qiuye Zhao Mitch Marcus Dept. of Computer & Information Science University of Pennsylvania qiuye, mitch@cis.upenn.edu Abstract We propose a bootstrapping algorithm which successfully resolves two funda- mental tasks: morphology acquisition and the acquisition of a subset of functional words. Given the outputs of these funda- mental tasks, we build a nearly state-of-art morphology analyzer performing with a F1-score of 80.94%; also, we can improve the baseline model for acquiring func- tional words by an absolute error reduction of 26%. Furthermore, with these acquisi- tion outputs, a minimally supervised tag- ging system proposed before can be turned into a totally unsupervised one, achieving a tagging accuracy of 85.26% for open- class words. 1 Introduction Studies of child language acquisition have shown that functional elements as distinctive categories are available to the child from very early on. These include both functional words (closed class items) such as determiners (Valian et al., 2009), and func- tional bound morphemes such as verbal inflec- tions (Yang, 2002). Motivated by such studies, we propose that functional categories, including both functional words and functional morphemes, should be identified first in the process of acquir- ing syntactic categories automatically from lan- guage input. Further motivation comes from the experimental results of the minimally supervised tagging system in (Zhao and Marcus, 2009)1, which, given only seven functional features, in- cluding four contextual features, whether modal 1The tagging system in (Zhao and Marcus, 2009) was re- ferred to as ’unsupervised’ in their paper, because back then, a lexicon was a common input to so-called unsupervised POS tagging systems. We classify their system as minimally su- pervised here, so as to differentiate it from this work which requires only raw text as input. verbs or determiners are left or right neighbours, and three specific morphological features, whether ’-ing’, ’-ed’, or ’-s’ are observed as endings, per- forms clustering in order to generate the two fun- damental open class categories, verbal vs. nomi- nal. This work suggests that functional elements are highly useful in further classification of open class items. This is quite different from most other POS in- duction systems in the Natural Language Process- ing (NLP) field for inducing Part-of-Speech (POS) tags, which, instead of generating clusters com- plying with the common understanding of ’syn- tactic’ categories, such as distinct clusters of de- terminers, nouns or inflections of verbs, tend to output scattered clusters consisting words of inter- esting similarities on many different dimensions (Christodoulopoulos et al., 2010). This is because, in these clustering-based systems, a vector space over the words is spanned by lexical features and suffix/prefix features, so that the generated clus- ters mix semantic and syntactic similarities (Clark, 2000). In this work, we explore an alternative ’top- down’ view of deriving categories, as opposed to the ’bottom-up’ view adopted by these earlier dis- tributional clustering methods. Here, we report …

Content truncated to 500 words