Slashdot Mirror
2006 Chatterbox Challenge In Full Swing
William Wynn writes "Once again chatterbots from around the world are coming together to face off in the ultimate bot competition. The 2006 Chatterbox Challenge lays host to 65 artificially intelligent programs attempting to imitate human conversation. Public voting takes place from April 1 to April 30 after which the private judging will have been finished and medals and cash prizes will be given out. Medals are awarded for "Most Popular Bot," "Best Learning Bot" and "Best New Bot" as well as $1,800 to be split among the top three bots overall. Anyone can talk to the competing chatbots through the competition website."
But they do know what they're talking about!
Elbot: "Who convinced you to come visit me?"
Me: "Slashdot"
Elbot: "That explains all these geeks who have been visiting me."
See!
Very little serious work has gone into chatbots since the 80s. The reason is that it's easy to make a chatbot that does practically nothing more than the textual equivalent of nodding whenever someone stops talking. It's hard to get any further with this approach. Actually chatting requires the bot to know something about some topic and be able to evaluate the other party's statements about it. Until both of these are done, you don't get a meaningful improvement. Furthermore, the system needs to be able to introspect its understanding of the topic. (There have been very old chatbots which actually did pretty well, by having a limited understanding of language and of some restricted topic. But these are relatively unintersting to chat with, compared to bots that let the user talk about anything and nod). Most of the recent AI research has involved making systems which behave appropriately in complex situations, and introspection isn't helpful for this (in people, this sort of stuff is preconscious, too; you can't explain how you recognize faces or voices or exactly how you identify spoken words). The things people have been working on turn out to have more direct practical uses, but they don't give the system anything to talk about. And, of course, there's been relatively little work on understanding arbitrary language since it became clear that it isn't that effective a way of communicating with computers anyway, because the human output side is slow and ambiguous, compared to other user interfaces.
Most likely, introspection in AI systems will be driven at some point by the need to combine different types of input to make a complete analysis (once there is sufficient success at handling those sorts of input), and language use will be driven by the need to handle language written for people as input (e.g., reading news reports for background information). At that point, it'll become reasonable to write an effective chatbot which talks about stuff that people care about.
I don't get it. Compared to Slashdot, that conversation was quite meaningful.
1) Who are you?
Slashdot: MICRO$OFT IS TEH SUXXOR
2) What are you planning to do today?
Slashdot: IMNSHO, STOP FEEDING THE TROLLS
3) What kinds of things do you know?
Slashdot: I FOR ONE WELCOME OUR COBOL-WRITING OVERLOARDS
4) What season is it?
Slashdot: WHEN I WAS 3, I WROTE MY FIRST LISP-COMPILER BY MAKING THE HOLES IN THE PUNCHCARDS WITH MY GUMS ONLY
Judge: Nevermind
Slashdot: LEARN TO TYPE PROPELY, YOU PSYCHO!
5) Where do you call home?
Slashdot: IANAL, BUT FIRST POST!