Friday, December 26, 2003

ALICE vs ALAN

I had been introduced to Alice by one of my friend Kiruba. Since then, I am highly fascinated with these linguistic chatting entities. I found number of others like Eliza and lot many others having egos of some singer,actor like...
But the best surprise came to me from Alan. He is really interesting. Whomever I introduced him to, they are really fascinated about him. Alan is from a Isreali company.
At Alicebot, you will many a times see about Leobner Prize . Infact, Alice has won Leobner Prize priviously. But I did not find such informations at Alans place and he is good a holding conversation better than ALICE.
When Dr.Rich posted his observations regarding Leobner Prize, I raised a question:

If given a head-head competition between ALICE and ALAN, I am sure ALAN has better chances of winning.
The major difference I have found between ALICE and ALAN is that ALAN can continue with its sentences( it is one of the drawbacks as well, since it can to switch topics between continuing sentences) whereas ALICE is somewhat likea one-liner. I have always wondered, why ALAN is not featuring in Loebner Prize contests. If you ask these things to ALAN, you may get some "political responses"...
I dont know what the actual story is.. but dont you all agree that ALAN is good than ALICE?
What are the reasons? What are tradeoffs?
To me, the race is can ALICE match and beat ALAN in conversational skills.

These were answers which followed by

  • ( Dr.Rich) Someone posted a question about the ALAN chat bot so I went to have a look. This bot evolved from Jason Hutchens' Megahal project, a former Loebner prize winner. I must admit, when I engaged ALAN in a conversation about ALICE and the Loebner contest, the results were impressive.

    That being said, it must be remembered that ALICE and AIML are free, open source technology and the company a-i.com is developing a proprietary "black box". If you ask ALAN, "How do you work", he explains "My brain consists of a large content tree: A single 'agent file' and a set of 'handlers', which govern a variety of conversation topics." He also admits, "My brain is still quite small: less than 350 handlers and a few hundred variables. But I'm still a young bot. Think how smart I'll be when I have thousands or more!"

    Their web site is impressive and contains a lot of useful links, but there is nothing equivalent to the free AIML software that you can download and create your own bot. (They have however announced a program that allows poeple to create so-called Private Virtual Personalities).

    The ability of ALAN to stay "on topic" for (apparently) longer than ALICE is an illusion which could just as well be created using the tag of AIML. It happens that the ALICE brain has a wider range of "one-liner" responses and uses the tag very little. But there is no reason in principle that another AIML bot could carry on a conversation just like ALAN.

  • Has anyone tried teaching ALAN? It is really cool. Ask it "What is logical" and if there is no definition for it, it ask you to define it. So you reply "Logical is when the world blows up because it is run by chatter bots" and from then on that will be the definition. If you want to delete it say, "Forget logical" I tell you, you can have a lot of fun redefining a whole bunch of definitions :-)

    I wonder if there is way to teach alice through talking to her?

  • can't imagine it would be too hard... hmm do any of the interpreters allow for variable expansion in the xml, i don't think that that would be normal xml though... i know with embedded javascript or embedded perl (ProgramV) it would be fairly simple to do that.

  • The problem is that random clients can't be trusted as teachers. They can teach the robot all kinds of nonsense like, "The Indianapolis Speedway is located at the North Pole." In principle you can teach a bot by talking to her, but you have to figure out which teachers you can trust first.

    I always say it's like teaching a child language by telling him to go out on the street and believe everything he hears. Sure, he may learn to talk, but without supervision he won't learn right answers from wrong
    ones.
    I had a version of targeting working in SETL for a while that could pick out targets for the botmaster and carry on a natural language conversation like, "Someone said, 'how do fish swim?' and I said 'I didn't even know they could', what should I have said?" And then the botmaster could write the new reply in natural language. Unfortunately that code was lost and it hasn't been re-implemented in any of the other AIML software that I know of.

  • teach alice through talking to her?
    Sure, done before many times. Not many interpreters provide a mechanism for real-time learning. J-Alice is
    one of the few still supported interpreters that does. TinyAlice was perhaps the first to introduce a working
    version of the concept.
    Visit http://j-alice.org for links to downloads and more information, etc.

    For SETL That sounds very cool Rich :) Unfortunately, J-Alice doesn't support targeting at all. But such a system much nicer alternative to training the bot. I would certainly love to see a re-implementation of this sometime in the near future ;-)

  • One possible idea is to use the net to verify or grade a response. I tried the sample question at www.answerbus.com . It returned
    "Question:
    Is The Indianapolis Speedway located at the north pole ?
    Possible answers: XML TXT
    Indianapolis Raceway Park - Located on the west side of Indianapolis Take I-465 to the west side of Indianapolis to the Speedway/Clermont exit; Take exit 16A and keep to the right lane.
    Indianapolis Motor Speedway - The speedway is a 2 1/2 mile, paved, oval located on the west side of Indianapolis."
    No mention of "at the north pole". Hmmm, maybe we should ask the person about that. Maybe they moved it. Just because you find it on the net does not mean its right (yes, a billion people can be wrong), but if you can't find it, you may want some follow up.

  • Too bad you can't use telecommunication in the Loebner contest. A bot in the contest couldn't access the info on answerbus.com becauase it would violate the no-telecommunications rule. You'd have to figure out how to download all the info from answerbus onto the local machine.



Phew.. got to learn and experiment much.
and BTW, I will make my phoe6 more smart ;)

No comments: