Regularized Interlingual Projections: Evaluation on Multilingual Transliteration

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 12–23, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Regularized Interlingual Projections: Evaluation on Multilingual Transliteration Jagadeesh Jagarlamudi University of Maryland College Park, USA, 20742 jags@umiacs.umd.edu Hal Daumé III University of Maryland College Park, USA, 20742 hal@umiacs.umd.edu Abstract In this paper, we address the problem of build- ing a multilingual transliteration system using an interlingual representation. Our approach uses international phonetic alphabet (IPA) to learn the interlingual representation and thus allows us to use any word and its IPA repre- sentation as a training example. Thus, our ap- proach requires only monolingual resources: a phoneme dictionary that lists words and their IPA representations.1 By adding a phoneme dictionary of a new language, we can readily build a transliteration system into any of the existing previous languages, without the ex- pense of all-pairs data or computation. We also propose a regularization framework for learning the interlingual representation, which accounts for language specific phonemic vari- ability, and thus it can find better mappings between languages. Experimental results on the name transliteration task in five diverse languages show a maximum improvement of 29% accuracy and an average improvement of 17% accuracy compared to a state-of-the-art baseline system. 1 Introduction Because of the wide usage of English, many natu- ral language processing (NLP) tasks have bilingual resources from English into other languages. For ex- ample, significantly larger parallel texts are available 1It is arguable that getting words and their IPA representa- tion require knowledge about both words and IPA symbols, but it still is specific to one language and, in this sense, we refer to it as a monolingual resource. between English and other languages. Similarly, bilingual dictionaries and transliteration data sets are more accessible from a language into English than into a different language. This situation has caused the NLP community to develop approaches which use a resource rich language (Q say English) as pivot to build resources/applications between a new lan- guage pair P and R. Previous studies in machine translation (Utiyama and Isahara, 2007; Paul and Sumita, 2011), transliteration (Khapra et al., 2010), and dictionary mining (Saralegi et al., 2011) show that these bridge language approaches perform com- petitively with approaches that use resources be- tween P and R. In this paper, we propose a regular- ization framework for bridge language approaches and show its effectiveness for name transliteration task. The key idea of our approach is that it accounts for language specific variation in the bridge lan- guage resources (i.e. between P ↔ Q and Q ↔ R) and aims to minimize this variation as much as pos- sible. Though our technique is general, for clarity we describe it in the context of named entity (NE) transliteration. Named entity (NE) transliteration involves transliterating a name in one language into another language and is shown to be crucial for machine translation (MT) (Knight and Graehl, 1998; Al- Onaizan and …

Content truncated to 500 words