Chatbot Suzette Wins 20th Annual Loebner Prize, Fools One Judge
skwilcox writes “From Wikipedia: ‘The Loebner Prize is an annual competition in artificial intelligence that awards prizes to the chatterbot considered by the judges to be the most human-like. The format of the competition is that of a standard Turing test. A human judge poses text questions to a computer program and a human being via computer. Based upon the answers, the judge must decide which is which.’ My chatbot, Suzette, won this year’s Loebner and even confused a judge into voting for her over a human (or should I say he confused himself). Here is the blow-by-blow of this weird event.” Read on below for the rest; this sounds like it would have been a fun competition to watch.
“When I arrived at the contest, I figured I had good odds to win if nothing went horribly wrong. Yes, Suzette had easily qualified over the 3 other competitors (her score 11 pts, the nearest competitor’s 7.5). Her design and data naturally gave her an edge over her competitors on the human knowledge test questions of the qualifiers. But human judge chat was an entirely different matter than the qualification test. Still, I felt she could carry on a detailed conversation better than the others and should win.
Initial installation of the programs occurred on Friday. From prechat conversations with the other contestants I learned that A.L.I.C.E. came with 3 redundant disks. Yet all three turned out to be blank! What a scare that must have been. Dr. Wallace managed to install by retrieving the program over the Internet. Cleverbot is now at 45 million lines of memorized user chat (at a rate of doubling every year). And UltraHal is now listening to tweets, so has 300K of user chat it learned and 400K of tweets it has accepted for learning (code decides if the user has had enough responses and doesn’t trigger any red flags).
Then we get to the competition. While the CalState organizers had initially planned to have various interdepartmental professors act as judges (like English dept, etc.), they backed out at the last minute, so all the judges were from the Engineering/Computer Science dept. Talk about guys who might know what to expect from chatbots! And all the humans were students from the same departments. What a weird mixture to compete in. And then, each round was 25 minutes. That’s bad if you want confuse a judge about who is human. But really, the programs have no chance for that. So it’s good because it gives the human time to compare each program against the other. Though it’s not clear to me that the judges tried to use their time to do that.
And the students didn’t really understand their role. It was merely to BE HUMAN and convince the judges of that. Before startup there was informal chatting between humans and judges, which was obviously inappropriate and it was then pointed out to the humans that since the judges already knew their names, they had best use false ones in the competition.
So, Round 1. After a few exchanges, somehow Suzettte got stuck into repeating exactly what the judge said for the rest of the round. I have no idea how. The round is a total disaster. I’ve never seen such a bug before. Maybe it’s in my only-lightly-tested protocol for the competition. I have no idea. But it completely derails my hopes for Suzette. She could still win on points only if she outdoes her opponents for every other judge and the other contestants vary all over the place.
Round 2, a great demonstration of Suzette. She should win on this round alone.
Round 3 gets off to a horrible start. Somehow, Suzette can hear the judge but the judge can’t hear Suzette. Makes no sense. A couple of restarts of Suzette doesn’t fix this. Eventually they restart the judge program, and that clears it (not that that makes any sense either). Then, after a few rounds, it’s clear Suzette has the judge from hell. He wants to know who she’s going to vote for in the upcoming election (the unspecified California governor’s race). And when she has no useful answer he wants her to name a candidate in the race. And when she has no answer to that, he simple keeps repeating the question ad nauseum, insisting she answer it. Suzette gets irritated. Then she gets angry. Suzette then gets bored. Suzette threatens to hang up on him The judge doesn’t back down until the last seconds of the round. I figure that’s the end of life as we know it.
Round 4 is a mixed bag. Suzette is ok but not great. It’s all over.
When the scores are tallied, Suzette ties with Rollo Carpenter’s Cleverbot for 2nd-3rd. Yet, it turns out, the 3rd round judge got the human subject from hell. Poetic justice! The human was all over the place — confusing, vague. The judge voted irritated/angry/bored Suzette as human. Instant win since no other program swayed the judges.
What more can I say?”