Enriching a statistical machine translation system trained on small parallel corpora with rule-based bilingual phrases