Analogical translation of unknown words in a statistical machine translation framework.
Etienne Denoual · 2007
In this paper we address the problem of translating unknown words in a statistical machine translation framework. In data-driven machine translation, words that are not seen in the data may not be translated and are either discarded or left as is in the output. They are refered to as unknown words. The unknown word problem increases when the available bilingual data is scarce. In order to address this problem, we propose to use proportional analogy at the character-level to translate unknown words. We study and report results of the integration of this approach into a statistical machine translation system translating from Japanese to English with relatively scarce resources. Objective evaluation measures suggest that the translated sentences have a higher adequacy than that produced by a baseline system, while their fluency is similar.