I'd train a linear regression on the text features to approximate the box office returns. That would be a way to build a rating system - the value of the function is the rating itself.