Learning from Multiple Sources for Data-to-Text and Text-to-Data