Improving transliteration with precise alignment of phoneme chunks and using contextual features

Abstract

Automatic transliteration of foreign names is basically regarded as a diminutive clone of the machine translation (MT) problem. It thus follows IBM's conventional MT models under the source-channel framework. Nonetheless, some parameters of this model dealing with zero-fertility words in the target sequences, can negatively impact transliteration effectiveness because of the inevitable inverted conditional probability estimation. Instead of source-channel, this paper presents a direct probabilistic transliteration model using contextual features of phonemes with a tailored alignment scheme for phoneme chunks. Experiments demonstrate superior performance over the source-channel for the task of English-Chinese transliteration. © Springer-Verlag Berlin Heidelberg 2005.