Finding translations for low-frequency words in comparable corpora