Video Recordings

The goal of dialogue system researchers has always been to create artefacts that offer natural interaction capabilities and effortlessly communication by means that humans also use to communicate among themselves. Even though systems like ChatGPT are already very good in form and style, there are more things to natural dialogue system behaviour than these LLM-based agents are capable of. I believe that this requires additional control capabilities of a dialogue system. In this talk, I will motivate this with insights from an analysis of communication styles in dialogues. I will continue with work on learning an explicit dialogue control component through reinforcement learning by optimizing on the estimated user satisfaction and thus ultimately improving the perceived naturalness of the interaction. I will finish with arguing that this basic idea is still relevant in the age of LLMs.