Curate a transliteration corpus from transliteration/translation pairs

Transliteration of new named entity is important for information retrieval that crosses two or multiple language. Rule-based machine transliteration is not satisfactory, since different information sources have different standards for the transliteration. To build a statistic machine transliteration module, researchers have to curate a transliteration corpus for any given two languages of interest. Since a large amount of transliteration/translation pairs can be collected from the Web, a large transliteration-training corpus can be curated from these pairs. In this paper, we proposed a bi-directional approach to classify transliteration/translation pairs. Our approach combines both forward transliteration and backward transliteration to classify transliteration from translation. An experiment on English and Chinese transliteration is conducted.