English to Hungarian Morpheme-based Statistical Machine Translation System with Reordering Rules

Phrase-based statistical machine translation systems can generate translations of reasonable quality in the case of language pairs with similar structure and word order. However, if the languages are more distant from a grammatical point of view, the quality of translations is much behind the expectations, since the baseline translation system cannot cope with long distance reordering of words and the mapping of word internal grammatical structures. In our paper, we present a method that tries to overcome these problems in the case of English-Hungarian translation by applying reordering rules prior to the translation process and by creating morpheme-based and factored models. Although automatic evaluation scores do not reliably reflect the improvement in all cases, human evaluation of our systems shows that readability and accuracy of the translations were improved both by reordering and applying richer models.