Joint Generation of Transliterations from Multiple Representations

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 943–952, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Joint Generation of Transliterations from Multiple Representations Lei Yao and Grzegorz Kondrak Department of Computing Science University of Alberta Edmonton, AB, Canada {lyao1,gkondrak}@ualberta.ca Abstract Machine transliteration is often referred to as phonetic translation. We show that transliterations incorporate information from both spelling and pronunciation, and pro- pose an effective model for joint transliter- ation generation from both representations. We further generalize this model to include transliterations from other languages, and en- hance it with reranking and lexicon features. We demonstrate significant improvements in transliteration accuracy on several datasets. 1 Introduction Transliteration is the conversion of a text from one script to another. When a new name like Ey- jafjallaj¨okull appears in the news, it needs to be promptly transliterated into dozens of languages. Computer-generated transliterations can be more ac- curate than those created by humans (Sherif and Kondrak, 2007). When the names in question origi- nate from languages that use the same writing script as the target language, they are likely to be copied verbatim; however, their pronunciation may still be ambiguous. Existing transliterations and transcrip- tions can help in establishing the correct pronuncia- tion (Bhargava and Kondrak, 2012). Transliteration is often defined as phonetic trans- lation (Zhang et al., 2012). In the idealized model of Knight and Graehl (1997), a bilingual expert pro- nounces a name in the source language, modifies the pronunciation to fit the target language phonology, and writes it down using the orthographic rules of the target script. In practice, however, it may be dif- ficult to guess the correct pronunciation of an unfa- miliar name from the spelling. Phonetic-based models of transliteration tend to achieve suboptimal performance. Al-Onaizan and Knight (2002) report that a spelling-based model outperforms a phonetic-based model even when pro- nunciations are extracted from a pronunciation dic- tionary. This can be attributed to the importance of the source orthography in the transliteration pro- cess. For example, the initial letters of the Russian transliterations of the names Chicano ([tSIkAno]) and Chicago ([SIkAgo]) are identical, but different from Shilo ([SIlo]). The contrast is likely due to the id- iosyncratic spelling of Chicago. Typical transliteration systems learn direct ortho- graphic mapping between the source and the tar- get languages from parallel training sets of word pairs (Zhang et al., 2012). Their accuracy is lim- ited by the fact that the training data is likely to con- tain names originating from different languages that have different romanization rules. For example, the Russian transliterations of Jedi, Juan, Jenins, Jelto- qsan, and Jecheon all differ in their initial letters. In addition, because of inconsistent correspondences between letters and phonemes in some languages, the pronunciation of a word may be difficult to de- rive from its orthographic form. We believe that transliteration is not simply pho- netic translation, but rather a process that combines both phonetic and orthographic information. …

Content truncated to 500 words