Phoneme-based Statistical Transliteration of Foreign Names for OOV Problem

Given a source language term, machine transliteration is to automatically generate the phonetic equivalents in a target language. It is useful in many cross language applications. Recently, there are increasing concerns about automatic transliteration, especially with languages with significant distinctions in their phonetic representations, e.g. English and Chinese. Despite many cross-language applications in English/Chinese, machine transliteration between the two languages has not been studied comprehensively. Existing English-Chinese transliteration techniques are typically based on source-channel framework, e.g. IBM SMT model. The accuracy of this model is rather low. In this thesis, we propose to use a direct approach for English-to-Chinese transliteration. We propose two direct transliteration models: In the first model, we model the problem as direct phonetic mapping from English phonemes to a set of rudimentary Chinese phonetic symbols plus dynamically discovered mapping units from training process. An effective algorithm for alignment of phoneme chunks is presented. In the second model, contextual features of each phoneme are taken into consideration by means of Maximum Entropy formalism, and it is further refined with the precise alignment scheme based on phoneme chunks. We compared the direct approaches with the source-channel baseline implemented with the IBM SMT model, and showed that the second approach is significantly superior.