Training Neural Response Selection for Task-Oriented Dialogue Systems