Jason D. Williams
Microsoft Research
Biography
Jason D. Williams is with Microsoft Research in Redmond, Washington, USA. He has published more than 55 peer-reviewed papers on dialog systems and related areas in international conferences and journals, and has received five best paper/presentation awards for work on statistical approaches to dialog systems, including the use of POMDPs (partially observable Markov decision processes), reinforcement learning, turn-taking, and empirical user studies.
In 2012 he initiated the Dialog State Tracking Challenge series, in 2014 he shipped components of the first release of Microsoft Cortana, and in 2015 he launched Microsoft Language Understanding Intelligent Service (www.luis.ai).
He is Vice President of SIGDIAL, and an elected member of the IEEE Speech and Language Technical Committee (SLTC) in the area of spoken dialogue systems. Prior to Microsoft, Jason was with AT&T Labs Research 2006-2012, and also held several positions in industry building spoken dialog systems, including at Tellme Networks (now Microsoft) as Voice Application Development Manager.
Systems he has built over the past 15 years have conducted tens of millions of dialogs with real users.
End-to-end deep learning of task-oriented dialog systems
This talk will present a model for end-to-end learning of task-oriented dialog systems. The main component of the model is a recurrent neural network which maps from a sequence of raw words directly to a distribution over system actions.
As compared to a conventional dialog system, the recurrent neural network takes the place of language understanding, dialog state tracking, and dialog control/policy. In addition, the developer can provide software that expresses business rules and provides access to programmatic APIs, enabling the network to take actions in the real world on behalf of the user. The neural network can be optimized using supervised learning (SL), where a domain expert provides example dialogs which the network should imitate; or using reinforcement learning (RL), where the system improves by interacting directly with end users.