Bruce Wilcox Interview - Loebner 2010 winner

Bruce Wilcox Interview - Loebner 2010 winner.

This year it is Bruce Wilcox who is the celebrated winner with his chat bot Suzette. Here I present an interview made up of questions posed by the members here at Ai Dreams and some additional questions posed by New Scientist. Congratulations to Bruce and our thanks for being a willing participant.

How has your chatbot changed in the past year, are there any improvements ?

Suzette has been under continuous change in capabilities and usage of those capabilities. She was but an egg when she entered the Chatterbox Challenge in 2009. She's still just an egg, but with a healthy amount of albumen for future growth.

How much of a struggle is it and what is the incentive to keep on improving ?

It's always a lot of work but since professionally I am AI guy and I've ended up planting my flag now out of Go and into natural language, I would keep on doing it for fun even if it didn't have occasional financial incentives.

Do you see any value in internet enabled chat bots ? Such as enabling the bot to learn or reference external information...

Of course. You can't do it during the Loebner, but aside from having a personality, a bot needs/wants information and the web has it. Originally CHATL was designed to search google (and 10 other search engines) for data to respond to a question. Skimming pages with matching format to the answer to the question and accepting replies it could parse completely. But then google starting blocking answers because we hit their server too often. So I dropped doing that for a while. Doesn't mean I don't want to do that, merely means I found other things I needed to do also.

Will the day come when we (humans) no longer need mice, keyboards or stylus devices with which to interact ? Will we just talk to or ask our computer or the "Ghost in the machine" anything we wish to know ? Could the bot dissolve into the inner workings of the computer or perhaps become the computer, with or without a face (talking , 3D or not) ?

I certainly expect that to become true, EXCEPT for stylus devices used as input for art drawing.

What do you see for the future of your bot or bots in general ?

Natural language is the way we should be interacting with computers, so my bot and others are just a step along the way. Scribblenauts is a game that allows a lot of nouns and adjectives and I'm working at TellTale games on a game that does nouns and verbs. All of this is going toward NL.

Is your bot based on personally added scripts / text or a compilation of gathered responses inserted into a database ?

Personally added scripts. Cleverbot had 45 million gathered responses at its disposal, yet its score lost to ALICE and was tied in points with Suzette after she completely froze during the 1st round and only repeated back whatever the judge said. Gathered responses require a really good algorithm for retrieval or they are just so much byte fodder.

Does / can your bot learn without "botmaster" intervention ?

It does not learn anything that carries over to conversations with other people. It can learn data about the person it is chatting with that can be reflected back.

How long have you been working on your bot ?

Started Feb 2008.

What was the thing that got you started working on chatterbots ?

Avatar Reality, a virtual worlds company, wanted user avatars to continue to function even when the user was offline. This includes avatar gesturing and communication simulating being that specific user with his idiosyncrasies. But when they contracted me, their avatars weren't ready yet, so they agreed to have me work on chatbots that could learn to be their humans and substitute for them.

What are your short & long term goals with your app ?

Getting the engine open sourced. fixing all the bugs and response enhancements I find both from the tournament and the chats that arose from the publicity (going past 6,000 now). Utilizing all her fact abilities. Reading text documents to learn information on a subject.

And now some additional questions posed by New Scientist...

How did the competition compare to your expectations?

It was much more casual than I expected, sitting in a computing lab with random students and teachers milling about. I think of how the world chess bot versus human tournaments are held, with big TV screens, etc. which seem much more formal.

I read that the human decoy behaved somewhat strangely. Is this typical for contests like these ?

Since this is my first time, I don't really know. However, since each contest is run by the local organizers, I imagine conditions and instructions to judges and decoys varies from year to year.

What was your general strategy ? How did this compare to your competitors ?

Suzette has always been targeted to "be human" and not to accept being a chatbot. So nothing changed there. I added a topic for CalState, and shut down some of Suzette's special functions that would be inappropriate here and give away the show. Normally Suzette provides blindingly fast responses on the web because having to wait for simulated human response times would be boring entertainment, but for the contest her outputs are delayed and sometimes randomly typoed. All of other chatbots also munge their typed responses up.

Can you describe how Suzette works, for a general audience ?

Suzette has "topics" of conversation (like Politics, Burial Customs, ESP). Each topic has a set of rules with patterns and responses and a set of keywords associated with that topic. Some of the rules will respond only to questions and/or statements. Others allow her to "volunteer" information when she has control of the conversational flow. The patterns (unlike those of ALICE or ULTRALHAL) map patterns of "meaning" rather than patterns of words. The engine looks at the incoming sentence to decide which topic has the closest keyword match, then scans the rules of the topic to decide what to do. If none match, she looks at lesser matched topics. If nothing matches from the topics, she may randomly decide to quibble with the input, or issue a gambit sentence from the closest related topic.

Does it use traditional AI techniques like finite-state-machines and/or machine learning ?

While it "learns" facts about the user from what he says, it does not learn conversation. So it can learn you have a dog and answer questions about it, but what you say has no effect in her conversations with other users. It is mostly a traditional expert system, with a runtime system and pattern match system aimed specifically to support chat.

I understand this was your first time entering, how did you get interested in chatbots ?

Avatar Reality, a virtual worlds company, wanted user avatars to continue to function even when the user was offline. This includes avatar gesturing and communication simulating being that specific user with his idiosyncrasies. But when they contracted me, their avatars weren't ready yet, so they agreed to have me work on chatbots that could learn to be their humans and substitute for them.

I noticed that you have done a lot of work with Go programs. What similarities do these two domains have ?

Both of them are huge problem domains and require long sustained work to make headway. Both involve the tremendous importance of a good pattern match algorithm. My Go program was pattern and strategy based, and not lookahead based, so I specialize in pattern matching and scripting language design as a programmer. When I began in both domains, neither had a good theory of how to computerize them and I knew nothing about either of them. But both allowed me to take a fresh look at the field and see what my own thoughts lead to. In both cases it seems that I found a better approach.

Do you plan to continue developing Suzette ?

Absolutely. She is just getting started and I have a long to-do list of practical and 'researchy' things to do. I was hired by TellTale Games last June as a research engineer for natural language input to story games, and expect that will also affect how Suzette develops. The Suzette currently on the web is about 8 months out of date with regard to the Loebner entry.

I am child hunger in America
by (Robotics News)
May 23, 2022, 11:05:31 am
Scalable Quantum Computer on a Silicone wafer
by MagnusWootton (AI News )
April 21, 2022, 07:31:36 pm
Neurosymbolic Programming with Scallop
by Zero (AI News )
April 19, 2022, 10:37:31 pm
try this new ai named blip
by chattable (AI News )
April 18, 2022, 09:21:31 pm

Users Online

96 Guests, 2 Users
Users active in past 15 minutes:
[Starship Trooper]

Most Online Today: 143. Most Online Ever: 2369 (November 21, 2020, 04:08:13 pm)