Abstract
The paper introduces a hybrid approach to a very specific field in machine translation - the translation of closely related languages. It mentions previous experiments performed forclosely related Scandinavian, Slavic, Turkic and Romanic languagesand describes a novel method, a combination of a simple shallowparser of the source language (Czech) combined with a stochasticranker of (parts of) sentences in the target language (Slovak, Russian, Slovenian).
The ranker exploits a simple stochastic model ofthe target language and its main task is to choose the best translation among those provided by the system. The last section of thepaper presents results indicating better translation quality comparedto the existing MT system for Czech and Slovak and compares themwith the results obtained by the translation from Czech to Russian using the same system.