Lightly Supervised Transliteration for Machine Translation

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 433–441, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Lightly Supervised Transliteration for Machine Translation Amit Kirschenbaum Department of Computer Science University of Haifa 31905 Haifa, Israel akirsche@cs.haifa.ac.il Shuly Wintner Department of Computer Science University of Haifa 31905 Haifa, Israel shuly@cs.haifa.ac.il Abstract We present a Hebrew to English transliter- ation method in the context of a machine translation system. Our method uses ma- chine learning to determine which terms are to be transliterated rather than trans- lated. The training corpus for this purpose includes only positive examples, acquired semi-automatically. Our classifier reduces more than 38% of the errors made by a baseline method. The identified terms are then transliterated. We present an SMT- based transliteration model trained with a parallel corpus extracted from Wikipedia using a fairly simple method which re- quires minimal knowledge. The correct re- sult is produced in more than 76% of the cases, and in 92% of the instances it is one of the top-5 results. We also demonstrate a small improvement in the performance of a Hebrew-to-English MT system that uses our transliteration module. 1 Introduction Transliteration is the process of converting terms written in one language into their approximate spelling or phonetic equivalents in another lan- guage. Transliteration is defined for a pair of lan- guages, a source language and a target language. The two languages may differ in their script sys- tems and phonetic inventories. This paper ad- dresses transliteration from Hebrew to English as part of a machine translation system. Transliteration of terms from Hebrew into En- glish is a hard task, for the most part because of the differences in the phonological and orthographic systems of the two languages. On the one hand, there are cases where a Hebrew letter can be pro- nounced in multiple ways. For example, Hebrew �ב can be pronounced either as [b] or as [v]. On the other hand, two different Hebrew sounds can be mapped into the same English letter. For exam- ple, both �ת and �ט are in most cases mapped to [t]. A major difficulty stems from the fact that in the Hebrew orthography (like Arabic), words are rep- resented as sequences of consonants where vow- els are only partially and very inconsistently rep- resented. Even letters that are considered as rep- resenting vowels may sometimes represent conso- nants, specifically �ו[v]/[o]/[u] and �י[y]/[i]. As a result, the mapping between Hebrew orthography and phonology is highly ambiguous. Transliteration has acquired a growing inter- est recently, particularly in the field of Machine Translation (MT). It handles those terms where no translation would suffice or even exist. Failing to recognize such terms would result in poor perfor- mance of the translation system. In the context of an MT system, one has to first identify which terms should be transliterated rather than trans- lated, and then produce a proper transliteration for these terms. We address both tasks in this …

Content truncated to 500 words