[kmoser]

You're invited to have a short conversation with a chatbot I am working on. The more you talk to it, the more it will learn. You can visit my chatbot here: http://chatbot.kmoser.com/. Your feedback is welcome. Thanks in advance!

[ivan.moony]

Hi

Do you use AIML or something else?

[Freddy]

Doesn't seem like a regular AIML set if it is. But yes it is learning some things. It repeated some things back to me.

Obviously this is a new bot so we must be kind 

The thing when it gets a response is a bit jarring...the please wait thing I mean.

Also it's trying to tell me a knock knock joke but it does not work.

Anyway, good start, welcome to AiDreams, please keep us posted

[Data]

I quite like the idea of having a blurred avatar.

The bot doesn't seem to have a response for Hello yet.

   ME: Hello

   BOT: That is not funny

I guess it's getting there.

Welcome to Dreams. 

[kmoser]

Thanks for everybody's feedback so far. This bot does not use AIML. The algorithm is actually very simple: everything you write gets stored verbatim in the database, and every response from the bot is one of those phrases. There is no semantic parsing of any kind. So you are really talking with everybody who has ever conversed with the bot (including yourself), using those exact phrases.

For example, if it has asked "How are you?" to ten people so far, and seven of them have responded, "I'm doing great!" then the next time somebody asks "How are you?" it is most likely to respond with "I'm doing great!". And if you write "I'm doing great!" it will pick the response to that which it has heard the most often...ad infinitum.

The secret sauce is how it decides which phrase to respond with, based on similar phrases it has heard before. The algorithm it uses to decide how closely two phrases match (e.g. "I'm doing great!" vs "Doing great!" vs. "I'm great!") is a combination of several MySQL pattern matching functions. That's probably the weakest link in the chain so far. The other weak link is that it has no ability to distinguish between what you said vs. what somebody else said, or when you said it. In fact, part of its strength comes from the fact that everything it hears is part of its "conversation", regardless of who said it or when. It does track sessions and IP addresses so theoretically I could make use of that in the algorithm but that gets into a whole other level of sophistication I'm not quite ready for.

The bot's success is predicated on the idea that when it comes to regular human conversation, given any phrase there will be a relatively limited set of possible meaningful responses. And given a long enough conversation, eventually it will be able to pick a meaningful response from the phrases it has heard before. Of course this requires lots of bot training, but if a few thousand people had a short conversation with the bot, it would become pretty smart quite quickly--provided those people didn't all discuss wildly divergent topics.

[Freddy]

I like the approach and find it fascinating how the bot may develop using this method. Do you consider most of the work done now or are there other major parts to still code ?

We have a sister site : www.chatterbotcollection.com

Maybe at some point you could submit your bot and we can hopefully get you some more chatters

[Art]

I chatted with it for about 10-15 minutes and while polite, I tried some different answers and a few questions.

Once in a while it even came very close with producing correct answers.

Nicely done and hopefully you will continue to "fine-tune" it's development.

Thanks!

[Freddy]

It's half learnt the knock knock joke now, it's got as far as..

- Knock Knock
- Who's there ?

Can't think Of how to teach it the rest...

- Etch
- Etch Who ?

[kmoser]

I like the approach and find it fascinating how the bot may develop using this method. Do you consider most of the work done now or are there other major parts to still code ?

We have a sister site : www.chatterbotcollection.com

Maybe at some point you could submit your bot and we can hopefully get you some more chatters

I've taken about as far as I can but I will occasionally tweak the algorithm. So I would consider it in beta test mode. At some point when I'm happy with it I will submit it to your other site. Thanks for the suggestion!

It's half learnt the knock knock joke now, it's got as far as..

- Knock Knock
- Who's there ?

Can't think Of how to teach it the rest...

- Etch
- Etch Who ?

As the algorithm stands, it will never learn the rest of the joke. This is because it has no "memory" of the earlier part of its conversation, so each statement stands on its own. It has a great memory for one-line jokes ("Why did the chicken cross the road?") but not for jokes that require multiple interactions that lead to the final punchline.