Current social conversational agents are trained in a supervised learning scenario on fixed human-human conversations, which makes it impossible to incrementally learn from interaction with real human partners and and does not match the intended usage in human-machine dialogues. In my project, I address this problem by initializing a recurrent neural network (RNN) based conversational agent on using large human-human conversations corpora, then taking a reinforcement learning actor-critic approach to incrementally optimize the dialogue generation performance by keeping the agent communicating with real users through the Amazon Alexa platform. The performance of this user-data-driven approach is expected to outperform RNN based approaches trained on fixed corpora only and to be competitive with other reinforcement learning based approaches using automatic feedback. My study paves the way for enrolling real users into the improvement of social conversational agents, which makes the intelligent agent learning procedure closer to real human learning.
For more details, please check my literature review Deep Learning for Response Generation and Deep Reinforcement Learning for Conversational Agents and my Research Proposal. Please feel free to e-mail me for questions and further discussion.
Before the CDT, I got my bachelor's degree in Computer Science in 2012 from Capital University of Economics and Business, Beijing, China. After that, I worked in the nature language processing department of Baidu Inc. for four years. During the last two year of my Baidu career, I was involved in Baidu voice assistant program, a task based dialogue system; Baidu voice map program, a task based dialogue system preferred to navigation and localisation, and Baidu Duer, a chat based dialogue system. My current research interests are chatbot systm and dialogue system.
Apart from my research interests, I also like landscape photography. Please feel free to check my photographs.