The CONVERSE program was not intended to be based on any scientific research but on hunches about how to do it, while taking advantage of some recent methodological shifts in computational linguistics towards empiricism and the use of real data. The main hunch was derived directly from PARRY's impressiveness when compared with its passive contemporaries like ELIZA. (Weizenbaum, 1976) PARRY had something to say, just as people do, and did not simply react to what you said to it. It could be said to rest on the hunch that a sufficient condition for humanness in conversation may be what Searle  calls intentionality: the apparent desire to act and affect surroundings through the conversation, which is a strong version of what we are calling ``having something to say", since a computer program without prostheses can only create such effects through speech acts and not real acts on physical objects. The extension of this hunch as far as Turing test situations are concerned - i.e. fooling people that the system is human-is that if the computer can get to say enough, to keep control of the conversation, as it were, through being interesting or demanding so that the human plays along, then there is correspondingly less opportunity for the human interlocutor to ask questions or get responses to an unconstrained range of utterances that will show up the system for what it is. Naturally enough, this hunch must be tempered in practice since a system that will not listen at all, and which will not be diverted from its script no matter what is said, is again inevitably shown up. The hunch is simply that, and translatable as: be as active and controlling in the conversation as RACTER, PARRY's only real rival (as regards interestingness) over the last 30 years, worked on the principle of being so interesting and zany that many humans did not want to interrupt it so as to intrude new topics or demands of their own. Others were less charmed of course, but it was one very effective strategy for operating this key hunch, and one not involving clinical madness, as PARRY did.
The original features of CONVERSE are as follows:
The last takes advantage of recent trends in natural language processing: the use of very large resources in language processing and intermediate results obtained from such resources, like the dialogue patterns. It meant that CONVERSE was actually far larger than any previous Loebner entry, and that much of our effort had gone into making such resources rapidly available in a PC environment. So, although not based on specific research, CONVERSE was making far more use of the tools and methods of current language processing research than most such systems. Its slogan at this level was ``big data, small program" which is much more the current trend in language processing and artificial intelligence generally than the opposite slogan, one which had ruled for decades and seen all such simulations as forms of complex reasoning, rather than the assembly of a vast array of cases and data. CONVERSE, although, it has some of the spirit of PARRY, does in fact have data bases and learns and stores facts, which PARRY never did, and will allow us in the future to expand its explicit reasoning capacity. The weighting system would in principle allow great flexibility in the system and could be trained, as Connectionist and neural network systems are trained, to give the best value of the weightings in terms of actual performance. We will continue to investigate this, and whether weightings in fact provide a good model of conversation-as opposed to purely deterministic systems that, say, always answer a question in the same way when it is posed. In the end, as so often, this may turn out to be a question of the application desired: a computer companion might be more functionally appropriate if weighted, since we seem to like our companions, spouses and pets to be a little unpredictable, even fractious. On the other hand, a computer model functioning as a counselor or advisor in a heath care situation, advising on the risks of a certain operation or test, might well be more deterministic, always answering a question and always telling all it knew about a subject when asked.