From rule-based to statistical grammars: Continuous improvement of large-scale spoken dialog systems