
www.Usenet.com
| <-- __Chronological__ --> | <-- __Thread__ --> |
"Jean-Baptiste Hétier" <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>... > I'm making a programe which is supposed to talk with you as in a chat or > something like. I've written the algorithm, but it now needs a brain. > > Could you tell me the few first questions you would ask someone you meet. > Thanks ! well, for someone to open a conversation, they must have an intention that cannot be reasonably be fulfilled by their own action, and can possibly be fulfilled by the other agent. So you have to look at what the agent intends, and then how he makes a dialogue plan that fulfills the intention. Of course with no information, the intention could be almost anything. So to avoid an ai-complete situation, you need to put the agents in one of those situations where there aren't that many intentions - at the travel agent shop, or in the bank. There are lots of corpora available of humans making dialogue in these situations which might be useful to give you statistics over intentions and plans. As for social settings where there is no definite intention, things may be a bit harder, but then again, people tend to go for lowest common denominator intention recognition such as knowing what someone does for a job etc - all the large-grained facts that might foster future cooperative interactions. i'm not sure of corpora that cover these settings, but i think the switchboard corpus is more domain-free, more "conversational" than others. if what you're asking for is the distribution of openings, working with a corpus might be useful
| <-- __Chronological__ --> | <-- __Thread__ --> |