Reduction of Morpho-Syntactic Features in Statistical Machine Translation of Highly Inflective Language

We address the problem of statistical machine translation from highly inflective language to less inflective one. The characteristics of inflective languages are generally not taken into account by the statistical machine translation system. Existing translation systems often treat different inflected word forms of the same lemma as if they were independent of each other, although some interdependencies exist. On the other hand we know that if we reduce inflected word forms to common lemmas, some information is lost. It would be reasonable to eliminate only the variations in inflected word forms, which are not relevant for translation. Inflectional features of words are defined by morpho-syntactic descriptions (MSD) tags and we want reduce them. To do this the explicit knowledge about both languages (source and target language) is needed. The idea of the paper is to find the information-bearing MSDs in source language by data-driven approach. The task is performed by a global optimization algorithm, named Differential Evolution. The experiments were performed using freely available parallel English–Slovenian corpus SVEZ-IJS, which is lemmatized and annotated with MSD tags. The results show a promising direction toward optimal subset of morpho-syntactic features.