SOPHIE is a conversational agent that converses freely in natural language. Her utterances appear as synchronized text under a colour photo representing her supposed visual appearance. During each chat the agent learns new facts and builds a profile for use in later conversations with the same user. her conversational repertoire has been built using a C-coded scripting language known as "Patternscript", and an interpretive engine "Interlex".. A Java version is in alpha-testing.

The aim is to test a position concerning human conversation, namely that people do not think while they talk, but freely associate to context-relevant patterns in the incoming discourse. In this model, conversation rests on automatized skill acquired in years of daily practice, similar to that of a grandmaster playing lightning chess "reflexively". This variant of the chess game allows an average response time of some 5 seconds a move, about the same as that available in human chat for the formulation and delivery of a response. Under such conditions chess-masters rely upon large memory-intensive and context-indexed pattern-stores, with no time tor "thinking". They respond intuitively. Do similar laws govern human real-time conversation?

A first step is to test whether mechanisms of associative context-switching and context-dependent pattern-matching form a sufficient basis for sustaining human-level conversation over short periods of time, - first formulated as a goal for Machine Intelligence by Alan Turing. The level attained by 2001 is described in the paper "Return of the Imitation Game" (see list of publications ). Improvements since then have confirmed that the performance of the current implementation adequately sustains conversations of ten minutes or so, so long as the human chat partner does not adopt too tricky or adversarial a style

The next task will be the development of intelligent editors with which progressively to automate the acquisition and scripting of the large masses of everyday knowledge needed to sustain human-level conversations of longer duration that are also more able to deal with idiosyncratic or hostile inputs.

The work was initiated in the laboratory of Prof. Claude Sammut at UNSW, Australia some five years ago. He designed and coded the original prototype. Professor Michie and his co-workers Richard Wheeler and David Mason aim to field the latest version in a suitable venue early in 2006.

Prof. Donald Michie
tel: UK +44 (020) 7722 1097