Policy-Driven Neural Response Generation for Knowledge-Grounded Dialogue Systems