Victor…

I’m not a developer or even a code writer, I’m just an average person off the street who’s been toying with several Pandorabots (5 at the moment) for a few years.  I’ll never enter the Loebner Prize Competition, but I enjoy reading about it and viewing the results, which I usually find disappointing, by the way.  In that sense, I view myself as representing the general public when it comes to chatbots although, admittedly, I have more than a casual interest in the subject.

I found chatbots because I had tired of chat room drama and the assorted shenanigans that seem to consume that corner of the online world.  I sought out chatbots because I wanted to chat.  In that respect, again, I think I represent the average user.  I think people go to chat with bots to chat, not to gain knowledge for a research paper.

I have nothing against bots that play games, or perform mathematical gymnastics.  Search engines serve a purpose and most of them do it well.  As a human, I’m not overly concerned with spelling or grammar, but as a botmaster, I’m forced to draw the line because there’s one way to spell a word correctly and zillions of ways to misspell it.  I’m not a stickler for proper grammar, but I’m keenly aware that improper grammar “sounds” awkward to the reader’s internal ear, and it’s necessary for patterns that my bots can recognize.

I think we have to understand why people want to visit chatbots, and what it is that chatbots do best.  If they can play games too, or play music, or display images or rare artwork, all the better.  But if their language “sounds” bad, or if they aren’t responsive and interrupt the flow of conversation with a non sequitur, no one will want to chat with them a second time.

T.W. :

Your approach is sound.  I have no use for the Loebner Competition, my bot will serve a purpose, that is, it is application oriented research.  I like your comment about “and what it is that bots do best”.  The aim for my bot project is to have an interactive dialog to help a user locate information *OR* to deduce information.

Erwin Van Lun - Jan 23, 2011:

I believe that learning a computer English is comparable to learning a computer Chess.

LOL ! We’ll have to get Erwin a very simple chat bot to help him with his learning/teaching challenge.    A simple, 1 or 2 line regular expression bot should do.

This brings up an interesting point.  Chatbots, although progress is slow and they may very well make errors, after being told (or code-adjusted, whatever the case may be), will never make the same linguistic mistake twice… they have no human ego to impede their pursuit to grow and learn, and take advice!  Ah, what a refreshing change !

So what is the deal with memory issues and chatbots?  Every bot I’ve seen, even the ones that are considered the best seem not to be able to remember what is spoken in previous lines, even when a clear reference is made.  I could understand if an indefinite article is used to refer back, but to use a proper name or noun, it seems like this problem should have been solved by now.

I think some bots have such memory capabilities but the creators, for various reasons, disable those functions when letting random users chat online with the bot. But I agree, a lack of conversational memory makes chatbots not appealing to interact with for any extended period of time.

Every bot I’ve seen, even the ones that are considered the best seem not to be able to remember what is spoken in previous lines, even when a clear reference is made.

I don’t know how advanced these features need to be to please you, but take a look at some of the test-cases I have. I’d say there is memory there. It’s not ‘there’ yet, but progress is coming along nicely.
There are others here on the forum as well who have bots with memory capabilities which can be tested/downloaded right now. (I’m thinking of Robert Mitchell,...).

Forgot Robert’s link

One week from today’s my friends !!!!!!!!!!!!!!!!!

Go Watson!

I have a friend here at work that argues that if Watson has to be given the question via keyboard input, that its not counting.  I disagree.

But I wonder though, will someone enter the text, or will Watson get the question electronically ?

Or, perhaps , does it actually do the visual recognition of the question on screen?  Also, what about audio? some questions are “name that song” - I don’t think it has that ability.

Personally, I think the point is NLP, which is the first step - then add audio later.  Some people are so hard to please!!!!!!!!!!!!!!!

If I read correctly, all questions are sent as ascii chars, the output is also only text, the voice is a trick.

Does anyone know where one can watch Jeopardy in Germany? Perhaps online somewhere? (Some sites restrict content to IP addresses in the US. Sometimes I can get around that with VPN, but sometimes not for some reason…)

They’ll probably post it on youtube or something somewhere. Don’t forget the entire purpose of this thing: advertisement for IBM, so it’ll probably be everywhere on the net.

Ok, my fascination with Watson has just PLUMMETED see the you tube URL below

Watson itself cannot read a sentence and understand it on its own?  3,000 cores. 

Ok, I give them credit for the HUGE amount of data it has, and the data correlation, but come on, a real AI needs to understand a sentence, figure out the meaning, completely on its own.

URL :

http://www.youtube.com/watch?v=d_yXV22O6n4&feature=relmfu

about the 3:13 mark.

That won’t cut it for a chat bot that can learn new concepts via natural language.

Watson itself cannot read a sentence and understand it on its own?  3,000 cores.

Watch it again and pay attention to what is said after that statement.  I think you took it the wrong way.

Watson itself cannot read a page of text and understand it.  That’s where we come in, we write the algorithms that help it break a sentence apart ... . ..

Hum, so one would have to almost see how the data entry works.  If it has to be some how guided on which of those algorithms is the correct one for understanding the source material. 

And he’s not very clear at that point.

The Jeopardy puzzle itself it deals with own its own. 

The actual processes of adding to its knowledge base—*IF* that is completly UNSUPERVISED, then ok, I’m impressed

But if, a human has to look at the sentence, and say, ok, I will indicate to Watson in some way, help it along, to which algorithm to use that will give it understanding so it can update its KB, then not so impressed.

It could be completely unsupervised, but then I’m wondering why the comment, “Watson itself doesn’t know how to read a sentence and understand it”—to me, that implies it is not fully unsupervised.

Now he DOES say “we write the algorithms that take **a** sentence”—which DOES imply supervised learning.

So, just strange why he said “Watson itself doesn’t’ know how…..”.  Perhaps initially it did not know, 

OR perhaps Watson itself , not the 100’s of algorithms.  The 100’s of algorithms each, individually know, and Watson, (core Watson logic or whatever), chooses which one is correct.

Watson doesn’t record the input. It’s knowledge base is database filled with ‘existing’ data like wordnet, framenet,.... + probably lots of manually added/tuned stuff. Remember how they were selling it as an example of what there DB2 is capable of doing?  It can’t add to it’s knowledge base itself. At it’s core it’s just a fancy query language. Data retrieval.