Transliteration System Using Pair HMM with Weighted FSTs

Abstract

This paper presents a transliteration system based on pair Hidden Markov Model (pair HMM) training and Weighted Finite State Transducer (WFST) techniques. Parameters used by WFSTs for transliteration generation are learned from a pair HMM. Parameters from pair-HMM training on English-Russian data sets are found to give better transliteration quality than parameters trained for WFSTs for corresponding structures. Training a pair HMM on English vowel bigrams and standard bigrams for Cyrillic Romanization, and using a few transformation rules on generated Russian transliterations to test for context improves the system's transliteration quality.