Omar
Zia
Khan, Senior
Applied
Scientist
Microsoft
In this talk I will discuss the system architecture for Cortana, Microsoft's personal digital assistant, with a particular focus on its single-turn and multi-turn dialog capabilities. I will describe a ranking based approach for multi-domain multi-turn dialogs. I will show how this can be extended to incorporate additional information from a speech recognition system to determine the correct system response. I will also discuss how to handle cases where no correct system response is available. Finally, I will present a technique to automatically handle queries referring to a user's past interactions with the system by combining an information retrieval approach with a personal knowledge graph. For all these components, I will also present experimental results based on data collected from actual Cortana users over time and how to incorporate user behavior to improve the product.
Bio: Omar Zia Khan is a senior applied scientist at Microsoft focusing on applications of natural language processing and machine learning. He currently leads a team that is building enterprise knowledge graphs using unsupervised and semi-supervised techniques by leveraging various enterprise artifacts such as emails, documents, and calendar appointments.
Before this, he was a scientist on the team that launched Cortana, Microsoft's digital assistant, and built various intent recognition and dialog management capabilities for Cortana. Prior to that, Omar was a part of an early stage startup, In the Chat, where he built a team that extracted intent and actionable insights for financial and telecom companies from social media data in real-time.
Omar is a graduate of University of Waterloo, where he obtained a Ph.D. in machine learning in the area of reinforcement learning and an M.Math. in computer science in the area of distributed systems.