Phoneme strings based machine transliteration

Abstract

Transliteration is always used to translate source names with approximate equivalence of pronunciation into target language. Current direct orthographical mapping (DOM) approach does segmentation and alignment on the basis of the single syllable. However it is hard to break down English names into constituent parts according to their corresponding single Chinese characters. This document proposes an approach of segmentation and alignment on the unit of phoneme strings in transliteration between English and Chinese. To lessen the calculation of model training on whole corpus, we split the training data into several pools stochastically and each is used to train a model. The final results of transliteration are arranged according to the decoding probability of each model, called combined model. The combined machine transliteration system between English and Chinese performs remarkably well on the shared task of NEWS2011. © 2011 IEEE.