Scalable and Safe Remediation of Defective Actions in Self-Learning Conversational Systems