Discriminative Substring Decoding for Transliteration

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1066–1075, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Discriminative Substring Decoding for Transliteration Colin Cherry and Hisami Suzuki Microsoft Research One Microsoft Way Redmond, WA, 98052 {colinc,hisamis}@microsoft.com Abstract We present a discriminative substring de- coder for transliteration. This decoder extends recent approaches for discrimi- native character transduction by allow- ing for a list of known target-language words, an important resource for translit- eration. Our approach improves upon Sherif and Kondrak’s (2007b) state-of-the- art decoder, creating a 28.5% relative im- provement in transliteration accuracy on a Japanese katakana-to-English task. We also conduct a controlled comparison of two feature paradigms for discriminative training: indicators and hybrid generative features. Surprisingly, the generative hy- brid outperforms its purely discriminative counterpart, despite losing access to rich source-context features. Finally, we show that machine transliterations have a posi- tive impact on machine translation quality, improving human judgments by 0.5 on a 4-point scale. 1 Introduction Transliteration occurs when a word is borrowed into a language with a different character set. The word is transcribed into the new character set in such a way as to maintain rough phonetic correspondence; for example, the English word hip-hop becomes ������ [hippuhoppu], when transliterated into Japanese. A task fre- quently of interest to the NLP community is back- transliteration, where one seeks the original word, given the borrowed form. We investigate machine transliteration as a method to handle out-of-vocabulary items in a Japanese-to-English translation system. More often than not, this will correspond to back- transliteration. Our goal is to prevent the copy- ing or deletion of Japanese words when they are missing from our statistical machine translation (SMT) system’s translation tables. This can have a substantial impact on the quality of SMT output, transforming translations of questionable useful- ness, such as: Avoid using a ������ account.1 into the far more informative: Avoid using a Freemail account. Though the techniques we present here are language-independent, we focus this study on the task of Japanese katakana-to-English back- transliteration. Katakana is one of the four char- acter types used in the Japanese writing system (along with hiragana, kanji and Roman alpha- bet), consisting of about 50 syllabic characters. It is used primarily to spell foreign loanwords (e.g., ������ [chokoreeto] — chocolate), and names (e.g., ����� [kurinton] — Clin- ton). Therefore, katakana is a strong indicator that a Japanese word can be back-transliterated. However, katakana can also be used to spell sci- entific names of animals and plants (e.g., �� [kamo] — duck), onomatopoeic expressions (e.g., ������ [bashabasha] — splash) and for- eign origin words that are not transliterations (e.g., ���� [hochikisu] — stapler). These un- transliterable cases constitute about 10% of the katakana words in our data. We employ a discriminative substring decoder for machine transliteration. Following Sherif and Kondrak (2007b), the decoder operates on short source substrings, with each operation producing one or more target characters, as shown in Fig- ure 1. However, where previous approaches em- ploy …

Content truncated to 500 words