I always inputted the optional word with slight errors. A totally wrong answer would be easy to filter out, but one or two letters being wrong while it's still a proper word would hopefully be harder to get out of the training data.