Lightly supervised transliteration for machine translation

Abstract

We present a Hebrew to English transliteration method in the context of a machine translation system. Our method uses machine learning to determine which terms are to be transliterated rather than translated. 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 requires minimal knowledge. The correct result 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.