English-Korean Machine Transliteration by Combining Statistical Model and Web Search

Machine transliteration is an automatic method for translating source language words into phonetically equivalent target language ones. Many previous methods were devoted to translating the word that only traces phonological phenomena of the source language and the resulting showed good performance. However, there are a lot of names originated from not only the source language but also non-source languages. The existing methods fail in showing high accuracy when the names comes from the non-source language since they focus on names in source language. To deal with this problem, this paper describes a hybrid method which combines statistical model and web search for improving machine transliteration performance. The proposed method constructs a base system that stands on a statistical model to produce candidates, then expands candidates from web documents. With these candidates, it finds the most appropriate answer without any external resources. The experimental results present that the proposed method achieves higher performance than statistical model and web search respectively.