Learning to Map into a Universal POS Tagset

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1368–1378, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Learning to Map into a Universal POS Tagset Yuan Zhang, Roi Reichart, Regina Barzilay Massachusetts Institute of Technology {yuanzh, roiri, regina}@csail.mit.edu Amir Globerson The Hebrew University gamir@cs.huji.ac.il Abstract We present an automatic method for mapping language-specific part-of-speech tags to a set of universal tags. This unified representation plays a crucial role in cross-lingual syntactic transfer of multilingual dependency parsers. Until now, however, such conversion schemes have been created manually. Our central hy- pothesis is that a valid mapping yields POS annotations with coherent linguistic proper- ties which are consistent across source and target languages. We encode this intuition in an objective function that captures a range of distributional and typological characteris- tics of the derived mapping. Given the ex- ponential size of the mapping space, we pro- pose a novel method for optimizing over soft mappings, and use entropy regularization to drive those towards hard mappings. Our re- sults demonstrate that automatically induced mappings rival the quality of their manually designed counterparts when evaluated in the context of multilingual parsing.1 1 Introduction In this paper, we explore an automatic method for mapping language-specific part-of-speech tags to a universal tagset. In multilingual parsing, this uni- fied input representation is required for cross-lingual syntactic transfer. Specifically, the universal tagset annotations enable an unlexicalized parser to capi- talize on annotations from one language when learn- ing a model for another. 1The source code and data for the work presented in this paper is available at http://groups.csail.mit.edu/ rbg/code/unitag/emnlp2012 While the notion of a universal POS tagset is widely accepted, in practice it is hardly ever used for annotation of monolingual resources. In fact, available POS annotations are designed to capture language-specific idiosyncrasies and therefore are substantially more detailed than a coarse universal tagset. To reconcile these cross-lingual annotation differences, a number of mapping schemes have been proposed in the parsing community (Zeman and Resnik, 2008; Petrov et al., 2011; Naseem et al., 2010). In all of these cases, the conversion is performed manually and has to be repeated for each language and annotation scheme anew. Despite the apparent simplicity, deriving a map- ping is by no means easy, even for humans. In fact, the universal tagsets manually induced by Petrov et al. (2011) and by Naseem et al. (2010) disagree on 10% of the tags. An example of such discrep- ancy is the mapping of the Japanese tag “PVfin” to the universal tag “particle” according to one scheme, and to “verb” according to another. Moreover, the quality of this conversion has a direct implication on the parsing performance. In the Japanese example above, this difference in mapping yields a 6.7% dif- ference in parsing accuracy. The goal of our work is to induce the mapping for a new language, utilizing existing manually- constructed mappings as training data. The exist- ing mappings developed in …

Content truncated to 500 words