Bottom-up earley deduction for preference-driven natural language processing