Robust Transliteration Mining from Comparable Corpora with Bilingual Topic Models

International Joint Conference on Natural Language Processing, pages 261–269, Nagoya, Japan, 14-18 October 2013. Robust Transliteration Mining from Comparable Corpora with Bilingual Topic Models John Richardson† Toshiaki Nakazawa‡ Sadao Kurohashi† †Graduate School of Informatics, Kyoto University, Kyoto 606-8501 ‡Japan Science and Technology Agency, Kawaguchi-shi, Saitama 332-0012 john@nlp.ist.i.kyoto-u.ac.jp, nakazawa@pa.jst.jp, kuro@i.kyoto-u.ac.jp Abstract We present a high-precision, language- independent transliteration framework applicable to bilingual lexicon extrac- tion. Our approach is to employ a bilingual topic model to enhance the output of a state-of-the-art grapheme- based transliteration baseline. We demonstrate that this method is able to extract a high-quality bilingual lexi- con from a comparable corpus, and we extend the topic model to propose a so- lution to the out-of-domain problem. 1 Introduction A large, high-quality bilingual lexicon is of great utility to any dictionary-based system that processes bilingual data. The ability to automatically generate such a lexicon with- out relying on expensive training data or pre- existing lexical resources allows us to find translations for rare and unknown words with high efficiency. Transliteration1 is particularly important as new words are often created by importing words from other languages, especially En- glish. It would be an almost impossible task to create and maintain a dictionary of such words by hand, as new words appear rapidly, especially in online texts, and word usage can vary over time. In this paper we construct a language- independent transliteration framework. Our model builds on previous transliteration work, improving extraction and generation precision by including semantic as well as purely lexical features. The proposed model can be trained 1This paper considers both ‘transliteration’ (EN– XX) and ‘back-transliteration’ (XX–EN). For simplic- ity we refer to both tasks as ‘transliteration’. on comparable corpora, thereby not relying on expensive or often unavailable parallel data. The motivation behind the approach of com- bining lexical and semantic features is that these two components are largely independent, greatly improving the effectiveness of their combination. This is particularly important for word-sense disambiguation. For example, a purely lexical approach is not sufficient to transliterate the Japanese ソース (soosu), as it can mean either ‘sauce’ or ‘source’ depending on the context. 2 Previous Work Previous work has considered various methods for transliteration, ranging from simple edit distance and noisy-channel models (Brill et al., 2001) to conditional random fields (Ganesh et al., 2008) and finite state automata (Noeman and Madkour, 2010). We construct a base- line by modelling transliteration as a Phrase- Based Statistical Machine Translation (PB- SMT) task, a popular and well-studied ap- proach (Matthews, 2007; Hong et al., 2009; Antony et al., 2010). The vast majority of previous work on transliteration has considered only lexical fea- tures, for example spelling similarity and transliteration symbol mapping, however we build on the inspiration of Li et al. (2007) and later Hagiwara and Sekine (2012), who in- troduced semantic features to a transliteration model. Li et al. (2007) proposed the concept of ‘semantic transliteration’, which is the con- sideration of inherent semantic information in transliterations. Their example is the influ- …

Content truncated to 500 words