An Effective Strategy for Modeling Score Ordinality and Non-uniform Intervals in Automated Speaking Assessment