Bayes decision rules and confidence measures for statistical machine translation

Abstract

In this paper, we re-visit the foundations of the statistical approach to machine translation and study two forms of the Bayes decision rule: the common rule for minimizing the number of string errors and a novel rule for minimizing the number of symbol errors. The Bayes decision rule for minimizing the number of string errors is widely used, but its justification is rarely questioned. We study the relationship between the Bayes decision rule, the underlying error measure, and word confidence measures for machine translation. The derived confidence measures are tested on the output of a state-of-the-art statistical machine translation system. Experimental comparison with existing confidence measures is presented on a translation task consisting of technical manuals.