Embodiments of the present invention provide a system for implementing multi-turn dialogs. The system performs a method that includes receiving a series of user utterances, generating a series of responsive system utterances, and labeling the series of responsive system utterances to generate training data for training a dialog management policy. The labeling includes executing a reward function at each turn of a dialog, in which for each turn of the dialog the reward function is configured to output a reward value that is based at least in part on an accuracy of the responsive system utterance of the turn and on number of dialog turns elapsed.