Finite-state models for lexical reordering in spoken language translation

The problem of machine translation can be viewed as consisting of two phases: (a) lexical choice phase where appropriate target language lexical items (words or phrases) are chosen for each source language lexical item and (b) reordering phase where the chosen target language lexical items are reordered to produce a meaningful target language string. In earlier work we have shown that nite-state models for lexical choice can be learned from bilingual corpora [6]. In this paper, we focus on stochastic nite-state models for lexical reordering and describe an algorithm to learn them from bilingual corpora. We have developed a stochastic nite-state English-Japanese translation system by composing nitestate lexical choice and lexical reordering model. We have evaluated it using the string edit distance of the translated string from a given reference string. Using this metric, the English-Japanese translation system scored 70.9% on English speech transcriptions.