Automatic evaluation of machine translation,based on computing n-gram similaritybetween system output and human referencetranslations, has revolutionized thedevelopment of MT systems. We explorethe use of syntactic information, includingconstituent labels and head-modifierdependencies, in computing similarity betweenoutput and reference. Our resultsshow that adding syntactic informationto the evaluation metric improves bothsentence-level and corpus-level correlationwith human judgments.
展开▼