On the Use of Bert for Automated Essay Scoring: Joint Learning of Multi-Scale Essay Representation