Erwin Van Lun - Jan 14, 2011:

Would anyone like to help the guys from IBM? Volunteering probably… I could make a phone call on behalf of a few of us. I guess they will be interested…

Yes, I would!

Tell them my work with A.I. has certainly aired on the TV News, and has otherwise aired in prime-time on TV, cross-referenced published in articles, they can independently verify.

Computers are vastly superior to humans at recalling lists of facts and trivia, so I’m not surprised when one of them beats one of us at some task. Chess programs can now defeat even the strongest human chess masters. But, when one of them can overcome the subtleties of language without detection, that will be truly impressive.

@Dave: Computers are stronger in chess since 1997. AI researchers are now working on general game play: learning a computer the rules of a new game, a let them beat human right away, without adding any game-specific knowledge.

Erwin,
learning => teaching

I don’t think that we have different words in Dutch for ‘teaching’ and ‘learning’. We tend to say: I learn you something, or I learn something (that, or I can’t remember the Dutch word for teaching). So it’s an easy mistake, I think. The English of course, do things differently, like driving
Besides, if you think about it, ‘teaching’ is not really possible, you can only serve info, it’s the one absorbing the info who does the learning (my explanation for the Dutch (or my) lack for the word).

I’ve remembered the word again: ‘les geven’ It’s even 2 (oh man).

Perhaps true, but the discussion here is English, thus we should use the correct terms.

And you are right, Teaching does mean to serve information.

Learning is the act, as you say, of “absorbing the info”

To make it clear:

“I will teach the chatbot”—

        I will provide information to the bot, explain concepts to it.  The chatbot will then “absorb the info”... it will learn.

“I will learn the chatbot to…...”

      This does NOT mean to serve info to the bot… this would mean, *I* am going to learn the chatbot.. that is, I am going to learn about the chatbot, I am going to learn how the chatbot works.  So here the chatbot is not learning something, *I* am learning about the chatbot.

tx mates, I’m teaching

lol…, yep,  you’re

Teaching yourself to use teaching when required.

but not

Learning yourself to use teaching when required.

Or,

You’re learning to use teaching when required.

but not

Teaching to use teaching when required.

@Erwin.

I think computers can beat humans at just about any game or task where rules are involved or a simple list of questions and answers.  Understanding language is clearly another ball game.  The convoluted way questions are asked on Jeopardy, is probably a good test.  I recall a scene in the 1982 movie, “Sophie’s Choice,” where she’s trying to learn English and complains about all of the words that mean the same thing like “velocity, fast, quick, swift, speedy, hasty, flit, brisk, expeditious, accelerated, winged”.

The biggest obstacle I encounter with my bots is trying to cover all of the ways there are to say or ask the same thing.  No matter how hard I try, there’s always someone who comes up with a way I hadn’t yet thought of.

And then, there are those who create compound sentences, including more than one idea in a question, and non-English speakers whose language is constructed different from English.  They know the right words, but the order is not the same.  Then, there are slang terms applicable to one language or culture and not another, figures of speech, and the wrong usage… your and you’re… to, too, and too… and how not all punctuation is recognized, the way humans use it as an indicator to differentiate between statements and questions in written language.

Any form of AI that can respond to conversational language in a coherent fashion is what I find impressive and interesting, no matter how many trivia questions it can answer.

Very well put Thunder Walk!

But we wouldn’t say the AI fails if it doesn’t understand a new way of someone saying something would we?  I mean, would a human child learning language, understand every possible way things could be said.  Carrying a conversation, and learning via conversation, these new synonyms and sentence structures that mean the same thing, would be the ultimate goal of NLP.  I can imagine how difficult a young child would have in parsing and understanding these sloppy and confusing English I see on some people’s facebook walls!

The way I am handing all these complexities in my design is with what I’m calling “permutation wrappers”.  I will first complete a bot that can understand language and expression, that can derive a common meaning regardless of whether any synonyms are used and regardless of what sentence structure is used.  Then, another permutation wrapper will, if it fails to derive the meaning, will start replacing different words, it could try replacing common prepositions, perhaps replacing ‘of’ with ‘from’, test again, ok, does the input generate a parse tree that makes sense now? Yes, go to next stage of processing, No, try another permutation wrapper, which may be, trying different punctuation marks, still doesn’t make sense, ok, next permutation wrapper is misspelled words.

Another permutation wrapper I have to work on is the apostrophe problem.

X’s Y

Should that expand to:

a) X is Y - example, if X=John Y=Smart,  so X’s Y = John is smart.

OR

b) stay as is, example X=John, Y=Car, so X’s Y = John’s car

But we don’t want “John’s car” to expand to “John is car”.

So basically I am working on a CORE functionality, where yes, you will have to use proper English, proper words (teach / learn),  correct spelling, etc.

THEN, layer by layer, add on the “pre-processing” stages to run through the algorithm multiple times with different words replaced (‘learning’ does generate any meaningful parse trees, ok, people often incorrectly use ‘learning’ instead).

The CPU, computing power you will require will depend on how tolerant your system is to incorrect usage.  For my system, CLUES, I believe it will be fast enough even on single thread with proper English.  I will know within the next 1-3 years (when i have say 10 layers of permutation wrapper/pre-processing), what it will be like.  I’m not really too worried about it though, now that I have taken the time to convert the engine to C++, I can make use of multi-threading (Perl multi threading really let me down!).

As a last note, consider a comparison of a calculator with an NLP engine - it is like we want a calculator to, when I would enter say 10 + 10 and it would say 20, I’d say WRONG!! FAIL !!  I “misspelled” the first 10, I *mean* 100, you’re system didn’t know that, so it fails.  That is about how unfair NLP is !!  Or perhaps every single word in a sentence is misspelled or has some vernacular definition the system didn’t know about.

I say we don’t worry about those things, when first building a chat bot.  Have it learn, common, normal, proper English first.  Then add on the layers of pre-processing (permutation wrappers), one by one, and cross your fingers you have enough horse power to run it!! perhaps on 10,000 machines in cloud computing.  Watson for example has 3,000 cores!

Sorry if there are errors in the above, please point them out, but I don’t have the time for proof reading   Short day and I’m dieing to get back to coding!!

Damn it! I think my chatbots.org posting addiction is coming back !!

@Thunder Walk - I highly agree with you that no matter how many questions Watson can answer regarding Jeopardy doesn’t prove it will be able to be deployed as a truly intelligent system which can carry on a conversation and understand language to its fullest extend - but it really doesn’t exactly disprove it either.

I guess the question is, does it really understand the English text, or is it doing some extremely powerful pattern matching and data mining/correlation that just simulates it?  Perhaps simulation is enough when you have THAT much data and that much computing horse power.  Does a chess algorithm really understand that it is playing chess, does it matter.  I guess we will wait and see.

It will be very interesting to see how this system will develop, and will truly be the first machine to pass a Turing Test?  I would love to see into the future 5 years !!!

Thunder Walk - Jan 23, 2011:

I think computers can beat humans at just about any game or task where rules are involved or a simple list of questions and answers.  Understanding language is clearly another ball game.

I believe that learning a computer English is comparable to learning a computer Chess. Once the problem is solved, we’ll continue with a more abstract challenge on a higher level.

Chess -> General Game Play
English -> Ability to communicate

The latter means: independent of language, of time frame, expressions, senses. Everything a human (or even a species) uses to communicate is understand by computers, they are learning and able to communicate back. That’s a real challenge for the coming 40 years.

40 years seems like an awfully long time.  With exponential increase in tech, maybe 10 years?  In the last 40 years we went from the first ICs to iPhones(and their augmented abilities).  I just think it will be shorter term.  Maybe the Jeopardy thing will be a Sputnik moment.

When I was completing my AI thesis back in 1993, people of my group at the university had been working on speech technology for more than 10 years!!! Actually, that’s not precisely what they had been doing, they actually worked on developing theoretical models, developing maths & algorithms for speech technology. The technology itself was not really available.

That has drastically changed the last years, one of the reasons of the the take off of speech technology nowadays.