10 November 2000, DIP Colloquium, Kristiina Jokinen

10 November 2000, DIP Colloquium, Kristiina Jokinen
Speaker: Kristiina Jokinen
Title: Evolution of Communication and Human-Computer Interaction: Prospects and Thoughts of Learning Dialogue Systems
Date and Time: Friday 10th November 2000, 15.00-17.00
Location: Department of Philosophy, Nieuwe Doelenstraat 15, Amsterdam, MFR (ground flour)

Abstract:
Development in computing power and computer facilities, together with the development in natural language processing, has led to a new view of the nature of computers: instead of regarding computers as simple programmable machines, the metaphor talks about communicating agents, intelligent dialogue systems, learning, and evolution. However, selections from a menu, pre-recorded responses, or repeated requests to rephrase one's question may not quite support the view of a system as an intelligent agent. In this talk I will discuss requirements for ambitious 21-century dialogue systems, and argue that one of the biggest problems in making dialogue systems more flexible, robust and natural is the knowledge acquisition problem. The more complex tasks the system is expected to cope with, the more complex knowledge is needed. However, the solution is not only to increase the system's memory but rather, to equip the system with capabilities to learn through interaction. An obvious choice to tackle the knowledge acquisition bottleneck is to use various machine-learning methods. It has be shown that the application of learning techniques results in substantial increase in processing speed as well as in accuracy and robustness of large-scale language processing systems. I will present some results done at CELE on comparing various methods on linguistic benchmark problems, and discuss their effect in the context of a learning dialogue system. Finally, the learning approach has also brought in such cognitive science problems as the innateness of language ability and evolution of cognition. Even if we leave the philosophical problems of consciousness and thinking machines aside, the models for symbolic learning systems can benefit from the investigations into neural and brain-like computation: general architectures, learning algorithms and memory organization form the basis for domain-dependent representations and attentional state, from which conceptualisation and language learning emerge as a result of interactions of constraints on various levels. I will finish my talk by drafting these kind of new perspectives on the representation and acquisition of linguistic knowledge.

More information can be found on the DIP (Discourse Processing) homepage, or by contacting the DIP Colloquium organizing committee at DIP@hum.uva.nl