Universal Morphological Analysis using Structured Nearest Neighbor Prediction

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 322–332, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Universal Morphological Analysis using Structured Nearest Neighbor Prediction Young-Bum Kim University of Wisconsin-Madison ybkim@cs.wisc.edu João V. Graça L2F INESC-ID Lisboa, Portugal joao.graca@l2f.inesc-id.pt Benjamin Snyder University of Wisconsin-Madison bsnyder@cs.wisc.edu Abstract In this paper, we consider the problem of un- supervised morphological analysis from a new angle. Past work has endeavored to design un- supervised learning methods which explicitly or implicitly encode inductive biases appropri- ate to the task at hand. We propose instead to treat morphological analysis as a structured prediction problem, where languages with la- beled data serve as training examples for un- labeled languages, without the assumption of parallel data. We define a universal morpho- logical feature space in which every language and its morphological analysis reside. We de- velop a novel structured nearest neighbor pre- diction method which seeks to find the mor- phological analysis for each unlabeled lan- guage which lies as close as possible in the feature space to a training language. We ap- ply our model to eight inflecting languages, and induce nominal morphology with substan- tially higher accuracy than a traditional, MDL- based approach. Our analysis indicates that accuracy continues to improve substantially as the number of training languages increases. 1 Introduction Over the past several decades, researchers in the nat- ural language processing community have focused most of their efforts on developing text processing tools and techniques for English (Bender, 2009), a morphologically simple language. Recently, in- creasing attention has been paid to the wide variety of other languages of the world. Most of these lan- guages still pose severe difficulties, due to (i) their lack of annotated textual data, and (ii) the fact that they exhibit linguistic structure not found in English, and are thus not immediately susceptible to many traditional NLP techniques. Consider the example of nominal part-of-speech analysis. The Penn Treebank defines only four En- glish noun tags (Marcus et al., 1994), and as a re- sult, it is easy to treat the words bearing these tags as completely distinct word classes, with no inter- nal morphological structure. In contrast, a compara- ble tagset for Hungarian includes 154 distinct noun tags (Erjavec, 2004), reflecting Hungarian’s rich in- flectional morphology. When dealing with such lan- guages, treating words as atoms leads to severe data sparsity problems. Because annotated resources do not exist for most morphologically rich languages, prior research has focused on unsupervised methods, with a focus on developing appropriate inductive biases. However, inductive biases and declarative knowledge are no- toriously difficult to encode in well-founded models. Even putting aside this practical matter, a universally correct inductive bias, if there is one, is unlikely to be be discovered by a priori reasoning alone. In this paper, we argue that languages for which we have gold-standard morphological analyses can be used as effective guides for languages lacking such resources. In other words, instead of treating each language’s …

Content truncated to 500 words