Transliteration rules generation and named entities transliteration conducted by a finite-state machine

The work proposes a method of cross-lingual transliteration rules generation. The method is based on analysis of a training set. Process of generation is divided into two stages: the first is extraction of simple (initial) rules, the second one is generation of complicated rules. The first stage uses the new alignment algorithm described in the paper. The transliteration is conducted by a finite-state machine constructed from a system of transliteration rules.