Architecture and Adaptive Machinery for Human-Computer Conversation seminar

A Computational Architecture for Conversation_AKAStudy

Andy made a presentation on this seminar and the introduction to this is as follows:

***Peeking Into The Seminar***

This seminar was to study dissertations on inference methods and structuring that make human-computer interaction possible. This is about two dissertations from professor Eric Horvitz, a researcher at Microsoft, a leading company in computer conversation, and professor Tim Paek from Stanford. The first paper is about research on the Bayesian Receptionist, which functions through the front desk of the Microsoft Seattle office. It is based on the Bayesian Network which uses Bayesian Inference which utilizes information other than just statistical sampling. It makes inferences by transferring its way of understanding the purpose of the conversation into a structure with 3 layers. For example, if the Bayesian Receptionist judges that the user’s purpose is to find a shuttle bus, it automatically identifies whether he is alone or in a group and which way the bus is going. Additionally, it works on continuing the conversation asking whether the user needs a VIP or handicapped seat. Concepts that are used here are Natural Language Processing, Value of Information, and Decision-theoretic Control other than just Bayesian inference. The second paper contains the analysis on linguistic and visual information, and the errors occurring in the conversational channels and signals. It also includes strategies to control the whole conversation, make corrections automatically, manage possible threats. It has been a task for machine learning to analyze the conversation and compile it into a structure that computers can understand. It might require a lot of research not only in the field of computer science, but also in the field of psychology and statistics to predict where the conversation is going on and the cost of analyzing these, etc.