Numerous commercial machine translation packages are readily available. While they differ in price, features, speed, etc., all have one common characteristic:  relatively poor translation quality.  Regardless of the glorious description on the box, they still cannot adequately satisfy a wide range of general-purpose translation needs.  Since their implementation and resources like lexicon size and coverage differ, each package has its own strengths and weaknesses.  This property can be exploited so that several packages in parallel produce a composite "averaged" translation with possibly higher quality than the translation of any individual package.  To this end, a variation on n-modular redundancy is employed, an architecture popular in engineering and computer science for mission-critical applications, where fault tolerance and error correction are highly desirable.  This paper covers a closely related architecture that helps to compensate for a variety of common translation errors in a relatively language-independent way.

Back to the Conference Schedule