Optimizing interactive systems with data-driven objectives

Building interactive systems requires a lot of effort, and understanding what users want and designing corresponding optimization objectives are some of the critical components. The reliability of hand-crafted objectives strongly relies on the amount of domain knowledge incorporated in them. In the first part of this thesis, we explore how to optimize interactive systems without hand-crafting objectives in a more general setup. Our solution requires no domain knowledge and is thus even applicable when prior knowledge is absent. In the second part of the thesis, we utilize the idea of data-driven objectives for two types of interactive systems: open-domain dialogue systems and task-oriented dialogue systems. Besides exploring the promising usage scenarios of data-driven objectives, we also investigate the limitations and potential problems of current deep reinforcement learning-based solutions for dialogue policy learning in task-oriented dialogue systems in the last part of this thesis.