Context-Based Morphological Disambiguation with Random Fields

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 475–482, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Context-Based Morphological Disambiguation with Random Fields∗ Noah A. Smith and David A. Smith and Roy W. Tromble Department of Computer Science / Center for Language and Speech Processing Johns Hopkins University, Baltimore, MD 21218 USA {nasmith,dasmith,royt}@cs.jhu.edu Abstract Finite-state approaches have been highly successful at describ- ing the morphological processes of many languages. Such approaches have largely focused on modeling the phone- or character-level processes that generate candidate lexical types, rather than tokens in context. For the full analysis of words in context, disambiguation is also required (Hakkani-T¨ur et al., 2000; Hajiˇc et al., 2001). In this paper, we apply a novel source-channel model to the problem of morphological disam- biguation (segmentation into morphemes, lemmatization, and POS tagging) for concatenative, templatic, and inflectional lan- guages. The channel model exploits an existing morphological dictionary, constraining each word’s analysis to be linguistically valid. The source model is a factored, conditionally-estimated random field (Lafferty et al., 2001) that learns to disambiguate the full sentence by modeling local contexts. Compared with baseline state-of-the-art methods, our method achieves statisti- cally significant error rate reductions on Korean, Arabic, and Czech, for various training set sizes and accuracy measures. 1 Introduction One of the great successes in computational linguistics has been the construction of morphological analyzers for diverse languages. Such tools take in words and enu- merate the possible morphological analyses—typically a sequence of morphemes, perhaps part-of-speech tagged. They are often encoded as finite-state transducers (Ka- plan and Kay, 1981; Koskenniemi, 1983; Beesley and Karttunen, 2003). What such tools do not provide is a means to dis- ambiguate a word in context. For languages with com- plex morphological systems (inflective, agglutinative, and polysynthetic languages, for example), a word form may have many analyses. To pick the right one, we must consider the word’s context. This problem has been tackled using statistical sequence models for Turk- ish (Hakkani-T¨ur et al., 2000) and Czech (Hajiˇc et al., 2001); their approaches (and ours) are not unlike POS tagging, albeit with complex tags. ∗This work was supported by a Fannie and John Hertz Foundation Fellowship, a NSF Fellowship, and a NDSEG Fel- lowship (sponsored by ARO and DOD). The views expressed are not necessarily endorsed by sponsors. We thank Eric Gold- lust and Markus Dreyer for Dyna language support and Jason Eisner, David Yarowsky, and three anonymous reviewers for comments that improved the paper. We also thank Jan Hajiˇc and Pavel Krbec for sharing their Czech tagger. In this paper, we describe context-based models for morphological disambiguation that take full account of existing morphological dictionaries by estimating condi- tionally against only dictionary-accepted analyses of a sentence (§2). These models are an instance of condi- tional random fields (CRFs; Lafferty et al., 2001) and include overlapping features. Our applications include diverse disambiguation frameworks and we make use of linguistically-inspired features, such as local lemma de- pendencies and …

Content truncated to 500 words