My current research revolves around open-domain dialogue research. By open-domain I mean that there are essentially no boundaries to what the conversational system can talk about. This is opposed to task-oriented dialogue, where typically the aim is to perform a certain task, such as an airline reservation. Open-domain dialogue can happen in various settings and with different goals. Social conversations happen in a social setting, similar to chatting informally with a friend or an acquaintance. In the social setting, it is important to have a personality with opinions, observing and responding to social cues and having empathy for the other person. In informative conversations the goal is to talk about relevant pieces of world knowledge seamlessly. Social conversations often intersect with informative conversations because people talk about the world in social settings as well. My research focuses on both: social conversations and informative conversations.
I spent a year co-leading the Stanford’s Alexa Prize team, Chirpy Cardinal. We designed – from scratch – a social bot (for having social and informative conversations) and deployed it to people all over the United States and had hundreds of thousands of conversations. It was an amazing experience to truly understand what users want, think about the full NLP stack and design systems that worked – not just for the users but also for the developers. We came 2nd place in the finals and are the first team to open source our entire codebase!
Not only was it an engineering challenge, it unearthed many fundamental issues with how we think about social and informative conversations. That formed the basis of the research questions that I am tackling: How to make informative conversations sound human and Understanding User Initiative. More on this coming soon, but feel free to reach out to hear me explain in more detail.