English-Chinese Name Transliteration by Latent Analogy

Name transliteration plays an important role in many natural language processing applications. Numerous machine learning techniques which adopt a top-down strategy have been applied to this task. Those approaches require a highly accurate English and Chinese alignment data and tend to dismiss relevant contexts when whey have been seen too infrequently in the alignment data. In this paper, a novel bottom-up approach for English-Chinese name transliteration that allows us to carry out direct orthographical mapping between two languages is proposed. Our English-Chinese transliteration approach has three steps. We first make a neighborhood of locally relevant transliteration names by using a latent semantic analysis of the appropriate graphemic form. We then align those names in neighborhoods via locally optimal sequence alignment. Finally, the maximum likelihood estimate is computed for every position to obtain probable Chinese transliteration of the given English name. The experimental results confirm its effectiveness in English-Chinese name transliteration.