The use of a structural n-gram language model in generation-heavy hybrid machine translation

Abstract

This paper describes the use of a statistical structural N-gram model in the natural language generation component of a Spanish-English generation-heavy hybrid machine translation system. A structural N-gram model captures the relationship between words in a dependency representation without taking into account the overall structure at the phrase level. The model is used together with other components in the system for lexical and structural selection. An evaluation of the machine translation system shows that the use of structural N-grams decreases runtime by 60% with no loss in translation quality. © Springer-Verlag 2004.