Noise-Aware Character Alignment for Bootstrapping Statistical Machine Transliteration from Bilingual Corpora

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 204–209, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Noise-aware Character Alignment for Bootstrapping Statistical Machine Transliteration from Bilingual Corpora Katsuhito Sudoh∗† Shinsuke Mori‡ Masaaki Nagata∗ ∗NTT Communication Science Laboratories †Graduate School of Informatics, Kyoto University ‡Academic Center for Computing and Media Studies, Kyoto University sudoh.katsuhito@lab.ntt.co.jp Abstract This paper proposes a novel noise-aware char- acter alignment method for bootstrapping sta- tistical machine transliteration from automat- ically extracted phrase pairs. The model is an extension of a Bayesian many-to-many alignment method for distinguishing non- transliteration (noise) parts in phrase pairs. It worked effectively in the experiments of boot- strapping Japanese-to-English statistical ma- chine transliteration in patent domain using patent bilingual corpora. 1 Introduction Transliteration is used for providing translations for source language words that have no appropriate counterparts in target language, such as some tech- nical terms and named entities. Statistical machine transliteration (Knight and Graehl, 1998) is a tech- nology to solve it in a statistical manner. Bilin- gual dictionaries can be used to train its model, but many of their entries are actually translation but not transliteration. Such non-transliteration pairs hurt the transliteration model and should be eliminated beforehand. Sajjad et al. (2012) proposed a method to iden- tify such non-transliteration pairs, and applied it successfully to noisy word pairs obtained from au- tomatic word alignment on bilingual corpora. enables the statistical machine transliteration to beIt bootstrapped from bilingual corpora. This approach is beneficial because it does not require carefully- developed bilingual transliteration dictionaries and it can learn domain-specific transliteration patterns from bilingual corpora in the target domain. How- ever, their transliteration mining approach is sample- wise; that is, it makes a decision whether a bilingual phrase pair is transliteration or not. Suppose that a compound word in a language A is transliterated into two words in another language B. Their corre- spondence may not be fully identified by automatic word alignment and a wrong alignment between the compound word in A and only one component word in B is found. The sample-wise mining cannot make a correct decision of partial transliteration on the aligned candidate, and may introduces noise to the statistical transliteration model. This paper proposes a novel transliteration mining method for such partial transliterations. The method uses a noise-aware character alignment model that distinguish non-transliteration (noise) parts from transliteration (signal) parts. The model is an ex- tension of a Bayesian alignment model (Finch and Sumita, 2010) and can be trained by a sampling al- gorithm extended for a constraint on noise. Our experiments of Japanese-to-English transliteration achieved 16% relative error reduction in transliter- ation accuracy from the sample-wise method. The main contribution of this paper is two-fold: • we formulate alignment over string pairs with partial noise and present a solution with a noise-aware alignment model; • we proved its effectiveness by experiments with frequent unknown words in actual Japanese-to-English patent translation data. 204 2 Bayesian many-to-many alignment We briefly …

Content truncated to 500 words