Customizing Watson's language model