The statistical approach to MT started about twenty-five years ago and has now been widely ac- cepted as an alternative to the classical approach with manually designed rules. Among the attrac- tive properties of the statistical approach is its ca- pability to learn the translation models automati- cally from a (sufficiently) large amount of source- target sentence pairs. Thus the need for the manual design of suitable rules and for human interaction can be reduced dramatically when developing an MT system for a new application or language pair. The idea of hybrid MT is to combine the ad- vantages of both the rule-based and statistical ap- proaches. In practice, most statistical MT sys- tems make use of manually designed rules in or- der to improve the MT accuracy. We revisit the RWTH syst...

Click Here to View the Paper In XML Format