Increasing the Quality and Quantity of Source Language Data for Unsupervised Cross-Lingual POS Tagging

International Joint Conference on Natural Language Processing, pages 1243–1249, Nagoya, Japan, 14-18 October 2013. Increasing the quality and quantity of source language data for unsupervised cross-lingual POS tagging Long Duong1,2, Paul Cook2, Steven Bird2 and Pavel Pecina1 1Faculty of Mathematics and Physics, Charles University in Prague 2Department of Computing and Information Systems, University of Melbourne lduong@student.unimelb.edu.au, paulcook@unimelb.edu.au, sbird@unimelb.edu.au, pecina@ufal.mff.cuni.cz Abstract Bilingual corpora offer a promising bridge between resource-rich and resource-poor languages, enabling the development of natural language processing systems for the latter. English is often selected as the resource-rich language, but another choice might give better performance. In this pa- per, we consider the task of unsupervised cross-lingual POS tagging, and construct a model that predicts the best source lan- guage for a given target language. In ex- periments on 9 languages, this model im- proves on using a single fixed source lan- guage. We then show that further improve- ments can be made by combining informa- tion from multiple source languages. 1 Introduction Supervised part-of-speech (POS) taggers perform very well in cases where substantial manually- annotated data is available, as is the case for languages such as English, Portuguese, German, French and Arabic. For example, Petrov et al. (2012) built supervised POS taggers for 22 Eu- ropean languages using the TNT tagger (Brants, 2000), with an average accuracy of 95.2%. How- ever, creating annotated linguistic resources is expensive and time-consuming. Many widely- spoken languages, such as Vietnamese, Javanese, and Lahnda have little or no manually annotated data, making a supervised approach impossible. However, parallel texts are becoming increas- ingly available through sources such as multilin- gual websites and documents, and large archives of translation memory from books, news, etc. Moreover, the number of languages with paral- lel data is increasing. The era of English dom- inating one side of parallel texts is shifting to a far wider range of languages. Parallel data can be exploited to bridge languages, and to trans- fer annotated information from a highly-resourced source language to a lesser-resourced target lan- guage, to build unsupervised POS taggers (e.g., Das and Petrov, 2011; Duong et al., 2013). One issue in building such a tagger is choosing the source language. English is commonly used, because parallel data which has English on one side is often most readily available. However, the appropriate source language might depend on the target language. For example, Snyder et al. (2008) found that a better tagger for Slovene could be built by using data from Serbian – a closely re- lated language – than from English. Moreover, if parallel data for a target language with more than one source language is available, it might be possi- ble to exploit this additional information; however, this issue has not been explored to date. In this paper we build unsupervised POS tag- gers for 72 language pairs. We identify features based on monolingual and parallel corpora that we use to predict the best source language to build a tagger for a given target language. We show that …

Content truncated to 500 words