TITLE: Virtual Language Understanding for a Virtual Assistant
PRESENTER: Peter Wallis
AFFILIATION: Department of Computer Science and Software Engineering The University of Melbourne
DATE: Wednesday, 16th January 2002
TIME: 2:00pm - 3.00pm
PLACE: CSE K17 1st Floor Seminar Room
The idea of a virtual assistant is appealing, but would appear to require a full natural language processing (nlp) interface and all the reasoning behind it. At the moment such interfaces only exist in science fiction. After 12 years working on applied nlp, I was about to start working on virtual assistants. Why? In this talk I will attempt to convince you that a virtual assistant can be a very different kind of problem to past work on natural language interfaces, and that the way ahead is to concentrate on models of politeness and service provision rather than on improving the ability of a system to understand. The proposed approach uses a BDI architecture to directly model the reasoning of an expert language user. The Belief, Desire, and Intentions architecture was initially created for embedded systems and is often described as reasoning for action rather than knowledge. Using this approach the goal of the virtual assistant's language processing unit is to, like Wittgenstein's brick layers, figure out what to say or do next, rather than to decipher the meaning of what is said. Having a model of language processing is one thing, but a key issue with any application in nlp is knowledge acquisition. How to populate the model with sufficient common sense so the system can do what it is intended to do without being overly distracting. Corpus analysis has provided many interesting results over the last 10 years, but requires a huge corpus when used on interactive dialog. An alternative, again from the BDI agent co mmunity, is to simply interview an expert and ask them what they intended to achieve with a particular act, and when they might have done it differently. The talk goes on to describe some preliminary, but interesting, experiments at DSTO in which we used techniques from Cognitive Task Analysis to create a BDI model of a human assistant booking divisional cars.
School of Computer Science & Engineering, UNSW.