Structural Feature Selection For English-Korean Statistical Machine Translation

When aligning texts in very di erent languages such as Korean and English, structural features beyond word or phrase give useful information. In this paper, we present a method for selecting structural features of two languages, from which we construct a model that assigns the conditional probabilities to corresponding tag sequences in bilingual EnglishKorean corpora. For tag sequence mapping between two langauges, we rst de ne a structural feature function which represents statistical properties of empirical distribution of a set of training samples. The system, based on maximum entropy concept, selects only features that produce high increases in loglikelihood of training samples. These structurally mapped features are more informative knowledge for statistical machine translation between English and Korean. Also, the information can help to reduce the parameter space of statistical alignment by eliminating syntactically unlikely alignments.