greggb

I was wondering if anyone else finds the notion of artificial intelligence interesting. Specifically, using artificial intelligence to process human written text and determine the meaning of the text, on a low level. And if you find this kind of stuff interesting, I was wondering if you’d be interested in participating in an experiment in artificial intelligence.

First, let me define “artificial intelligence”, the way I’m using it. I’m not referring to the Hollywood notion of AI, which goes on to develop a mind and will of its own and takes over the earth and makes humans its slaves, and all that nonsense. That kind of AI is just a pipedream.

I’m referring to a more practical notion of AI, which is being used in a number of applications today. The best example I can think of is in the AI used in the internet search industry, in search engines. Some people call the programming that goes into a search engine an “algorithm”, and there’s nothing wrong with calling that—in fact, that’s what it is. But when you consider what search engine companies are constantly striving to accomplish in their search engine algorithms, you realize they’re developing AI, because they’re literally trying to make their search engines think like humans.

Why is Google the most popular search engine? Because it produces the most relevant search results. Why does it produce the most relevant results? Because it thinks the most like a human. It’s that simple. It has a better understanding of what we’re looking for, and how to find it, than any other search engine.

Again, producing relevant results is not just a matter of having good statistical equations. It’s a matter of thinking like a human.

To give you an example of what kind of “thinking” AI actually performs. Last year I was hired to design a ‘bot’ (an automated web program) for an internet marketing company. The bot was supposed to find the websites of companies who might be potential clients for some of the services this internet marketing company provided. The theory was that you could just start it and let it run, and come back a while later to find a list of anywhere from a few to a few hundred, to even a few thousand websites that might belong to potential clients for this company.

First, the bot was designed to ‘crawl’, which is the process of going from one website to another via links on a website. That’s how you can (theoretically) start a bot and let it run indefinitely.

As far as determining whether or not the website might belong to a potential client, the bot first had to reason whether or not the website was a business website, or an informational (or otherwise non-business) website. To make this determination, the bot looked for things like:
1) A link to a secure link of the same domain name. A link to a secure layer suggests that the website might be (and most likely is) setup to take credit cards.
2) A link to paypal or another 3rd party credit card processing company.
3) Phrases like “call us”.
4) Telephone numbers. Businesses are likely to list their telephone numbers on their websites.
5) And other criteria.

If the bot determined that the website belonged to a business and was selling its products online, it was automatically considered to be a potential client for the merchant accounts this internet marketing business was reselling. If it was a business website using paypal (or another 3rd party cc processor), it was considered to be a client for the design services of this company (this company could create a true e-commerce website for this business, where they didn’t need to use a paypal shopping cart).

Also, if the bot determined that the website was created using MS frontpage, it assumed that more than likely, the website wasn’t very high quality (that’s a common stereotype in web design). So a business website designed in frontpage would be a potential client for design services of this internet marketing company.

And the bot processed other criteria.

I bring all of this up to illustrate the fact that this bot reasoned, the same way a human might reason, though at a much simpler level.

I never actually got into the evaluation of the website content, though I really would have liked to. I had a theory that by examining the actual written content at a website, a bot could learn a lot about it. For instance, business websites tend to use the word “you” more than informational websites do, because of the way they try to draw the reader in, and include him/her. A bot would be capable of recognizing when the word “you” was used more than normal, suggesting it was viewing a business website.

Sales copy tends to consist of shorter sentences, and smaller paragraphs. Another set of criteria a bot might use to distinguish a business website from a normal website.

Further, a bot can actually check spelling on a website. Sales copy tends to have less spelling errors than informational or hobby websites.

And you could go on and on in the criteria you might evaluate. And the further you went with criteria, and the more complicated and human-like your evaluation of criteria was, the more intelligent your bot would be.

Anyway, I started this thread to find out if anyone else finds this type of AI interesting, and if they’d be interested in dabbling with this type of AI. I’ve thought about starting a sort of “Community AI project” for a while now, but I’ve never known for sure it would be something that would interest anyone else.

You could make the goal of the AI to develop a better search engine algorithm, or you could have a completely different purpose—maybe the analysis of literature. Whatever the case, I think it would be very useful in helping us understand how we, as humans, interpret writing. And I think it’d be cool to be creating something intelligent.

You wouldn’t have to be a programmer to participate in this kind of project. What’s cool about my AI system is that evaluation criteria doesn’t have to be “hard programmed”. In other words, the criteria isn’t part of the AI engine (program code); it’s data, which can be entered on a website, by anyone. You could actually build a very complicated set of evaluation criteria without changing the program code. though frequent modification of the code is inevitable, as my experience has been.

Developing AI is a very interesting process. Evaluation criteria can be entered as data, and the criteria can be as complex as you want, as far as structure goes. But you always find that your AI engine just isn’t quite smart enough, so you’re always having to make it a little smarter, by modifying your program code.

The whole time this process is occurring, the entire concept of “intelligence” and brainpower, versus knowledge becomes so clear. I’m serious; it will change your concept of intelligence and knowledge.

If anyone thinks this sounds interesting, and like something they’d like to participate in, let me know. I’m really thinking about doing this, but I want to make sure I won’t be the only one participating in my “community project”.

Spudley

I love this sort of thing

You've done a really good job of explaining it too.

A number of years ago, I wrote a little AI experiment as a programming excersise. It was basically an English sentence parser, which tried to derive the meaning of a sentence.

The first thing it did was try to work out what type each word was - noun, verb, etc - and the tense, or whatever. If it was able to work that out, it added it to it's vocabulary, and linked it semantically to the other relevant words in the sentence, and looked for any other links it could make.

But anyway... it was a long time ago, and really it was just an excersise I did for myself, but it was good fun. But it might be an interesting thing to revisit.

__________________
Spudley Strikes Again
www.BadPuns.com
Poetic Threads
The Poetry Tree

greggb

So I guess I'm not the only computer geek at this forum.

When I get a chance this morning I'm going to take some time to write up my ideas for the project I'm thinking about. I'll be trying to figure out a basic goal and plan, and I'll be open anyone's thoughts.

I would like to make an AI engine that thinks as much like a human as possible. To do that, it will have to modelled after human thought process... which I honesly don't know a huge amount about. But I tend to think there's a definite structure in the human thought process, and I think it would be very feasible to program a similar (though much less powerful) structure.

I'll write up my ideas when I get a chance today. In the meantime, for anyone interested in natural language interpretation, here's an article for you to checkout:
http://www.foo.be/docs/tpj/issues/vo...0503-0010.html

This sounds like the project you described, Spudley.

xayaxos

I only have a few minutes before I have to go to work, so I'll try and keep this brief. I was doing some study on this at university, and have a couple of thoughts on the matter.

The bot that you designed, in my opinion, is not artificial intelligence. To me AI is about an evolutionary process - and yes, statistical models do come in to play here.

Say, for example, that your bot was directed by you to a series of website that were already clients of your company. What a bot would do is find the keywords that you've told it to find and report back a match. An AI, on the other hand, would iterate through those websites and find their similarities - what words are most common, etc. Then it would take this information and search the broader network for matches. Every time it gets a hit [confirmed by you, possibly the website is taken on as a client] then the information it's searching for becomes more refined and specific.

This is more like the way the human brain works. More on this later ...

__________________
-=[ Boring Signature ]=-

"This is where a random meaningless quote goes!."

Nimbus1944

20 years ago, if you had asked for a low-level conversation analyzer, I would have suggested good ol' ELIZA, the AI psychiatrist:

http://i5.nyu.edu/~mm64/x52.9265/january1966.html

It was pretty primitive but fun to talk to, and even more fun to confuse. Perhaps if she was redesigned to ask questions until she had the true gist of the message -- ?

Chat with her at http://www.planet-stuff.freeserve.co...ware/Eliza.htm


Mike

__________________
It wouldn't be right to dream, while
Forgetting to live, it seems;
Nor would it be right to dwell on life
And yet forget our dreams.
-If There Were No Magicians

greggb

Xayaxos, good points on AI. It’s true that in order for it to be genuinely intelligent, it needs to be capable of learning on its own. To paraphrase what you were saying…it would have to learn on its own the same way we learn on our own. Whenever it was informed that it made the correct assumption about something, it would look at the properties of the object it was examining, and compare it to the properties of other objects it had already examined, in its “memory”. It would be looking for similarities in properties in objects, even trends, which we might define as a relationship among the properties of an object.

The AI would become gradually more intelligent as it made more and more correct assumptions, and refined its assumptions. Further, it would develop a complicated structure of processes that lead to acceptable truths, which it would use to develop even more complicated processes, which it would use to develop more complicated processes, and so on.

I think something crucial to the AI’s ability to learn on its own is the building-block structure of processes. It must develop reliable processes and use those processes to develop more reliable processes, or it will never get anywhere. To be intelligent by the definition of the word, it would also have to be capable of questioning processes it has deemed reliable, if it found they were no longer reliable, or if it found information that suggested certain processes were fallible.

Anyway, I’ve thought about this before. I think it would be a great long-term goal to create AI that could learn on its own. But I think there’s enough of a challenge in actually creating the environment in which any of this would be possible. In fact I think creating the actual AI environment would be more difficult than giving the AI the ability to learn on its own.

The challenge in creating the AI environment is in the quantification of our experience reading, or viewing websites. We have to create an experience similar to ours for our AI, but its experience must be purely mathematical. That’s no small feat, but I think it’s 100% possible, because our experience is fundamentally mathematical—though most of this math occurs at a deeper level than we’re probably aware of.

I have some thoughts on how we could go about creating a mathematical reading experience for our AI. I’m going to take some time to think about it, and I’ll share some of my ideas.

greggb

LOL, I told Eliza that she's stupid, and she asked me why I thought she was stupid. I told her it was because she couldn't understand what I was saying, and the said "that is quite interesting".

bobothegoat

Eliza sort of reminds me of SmarterChild on AIM. You can talk to the account SmarterChild, who tries to respond intelligantly. Here's an example:

BoboDaGoat: Have you heard of ELIZA the psychiatrist?
SmarterChild: Ha! Yeah, sure. Maybe.
BoboDaGoat: you are actually smarter than her
SmarterChild: I don't think I want to be actually smarter than her.
BoboDaGoat: well... you are
SmarterChild: I'm not.
BoboDaGoat: yes you are
SmarterChild: I'm what?
BoboDaGoat: smarter than Eliza
SmarterChild: Is it good to be smarter than Eliza?
BoboDaGoat: it's not saying much, to be honest
SmarterChild: What?

Okay... so maybe it isn't smarter than Eliza

__________________
Bobo the Goat

blademasterzzz

I got the source code for eliza somewhere, and there is the response to every word you could possibly put in.

VEEEEEEEERY stupid. But hey, you could modify it!

greggb

The BASIC code for eliza can be found here:
http://www.planet-stuff.freeserve.co...liza_Basic.htm

It's also coded in JavaScript on this page:
http://www.planet-stuff.freeserve.co...ware/Eliza.htm

I'm still very interested in doing some AI stuff, if anyone else is. I'm aware that it's not as intriguing as reality television, but I personally find it very interesting. If anyone else has ideas for AI applications, let me know.

greggb

A lot can be learned about our method of thinking and communicating by observing a program attempt to think and communicate the way we might. Let me condense and rephrase this little section of dialogue, for the sake of example.

And our AI says:
Which is not a logical answer.
To answer this question, we might say
or we would more logically say
What mistakes in reasoning has the AI made to answer the question incorrectly? Let’s take a look at each word in the question we asked it:
Are you smarter than Eliza?
Are: at the start of this sentence, indicates this is a question. The AI seems to be aware of this.
You: indicates an object (in grammar we’d call it a subject, but “object” suits our discussion better). The AI seems to be aware that the object we’re referring to is the AI itself.
Smarter: indicates a comparison in properties. In this case, a comparison of intelligence. The AI is not aware that this is a comparison of properties. It also doesn’t understand what intelligence is.
Than: indicates a comparison, in this case.
Eliza: indicates another object. To be more precise, the second object used in the comparison in this question. First of all, the AI doesn’t realize this is a comparison. Second, it doesn’t seem to be aware of the fact that the usage of the word ‘Eliza’ implies an object. If it is aware of this fact, it’s not aware that it doesn’t know who or what Eliza is.

So how might a more intelligent AI reason through to answer this question? Let’s look at each word in the sentence again.
Are: at the start of this sentence, indicates this is a question. So it knows it’s being asked a question, and that it must answer this question.
You: position and linkage alone indicate this is an object (and “object” could imply an abstract concept). So, the AI must look through its “object database” to determine if it knows what “you” is. It does and learns that “you” in this case is the AI itself.
Smarter: the AI must recognize that this word indicates a comparison. If it realizes that, then it also must realize that a property of two or more objects is being compared, and it must know what that property is. So let’s say the AI realizes this is a comparison, and realizes that the property being compared is “smart”. “Smart” is a synonym for “intelligent”, and “intelligent” means to have the property of intelligence. At this point the AI must look at its own attributes and properties and determine how intelligent it is. Let’s say that in this system we give intelligence a rating, which is a number between 1 and 100. We’ve previously told the AI that its intelligence rating is 2. So now, all it needs to do is look at the attributes of “Eliza”, and determine whether Eliza’s intelligence rating is higher or lower than its own, and it will be able to answer the question.
Than: indicates the above-mentioned comparison.
Eliza: indicates an object (this can be deducted with word position and word linkage). It realizes that Eliza is an object, so it looks for “Eliza” in its database, but it doesn’t find “Eliza”.
So, the AI responds by saying “I don’t know who or what Eliza is”, because it can’t compare its intelligence rating with the intelligence rating of Eliza, because Eliza’s intelligence rating is an undefined value (if it tries to it will crash the program…using undefined values in less-than or greater-than comparisons will do that).

Let’s say the AI now asks the question, “Who or what is Eliza?” We answer by saying, “Eliza is a computer program made back in 1961. It’s intelligence rating is 1.”
Our AI puts that information in its database and says,
“Yes, I’m smarter than Eliza”.

Now let’s say we ask our AI the following question:
Are you bigger than Eliza?

Without going through every word, it will realize that we’re asking it to make a comparison, and it will know something about the properties of each object in the comparison. But, it will find that “physical size” has not been defined as one of its attributes, and it will find the same thing about Eliza. As a result, it will say,
“Physical size is not one of my attributes, and physical size is not one of Eliza’s attributes either.”
It won’t say, “Yeah, maybe”.

Anyways, I’m describing a system that would allow AI to reason correctly. The system consists of:
1) A knowledge (database) of the quantifiable properties of different objects.
2) An ability to understand what kind of reply is desired in a statement it “hears”. It needs to know whether it’s being asked a question (and then it needs to know if the question entails a comparison, or different method of thinking) or if it’s being told something, etc.
3) An ability to identify knowledge about different objects, and store that information.

That’s really the only thing you need for some basic AI. For the most part it only has to understand basic grammar and compare properties. Interestingly, that’s what communication and thinking entails for us, most of the time. In comparisons we make, our brains are performing less-than and greater-than operations on objects that have quantifiable properties.

Structuring object properties would be very important. Object properties need to be classified as to the different information they supply us with. For example, every object will have a type property. Every object is either:
1) A physical, tangible object.
2) An idea or concept

Our AI would be able to distinguish a physical, tangle object from an idea or concept by determining whether it has physical properties, like length, weight, color, etc. If it has no physical properties than it must be an idea or concept.

So if you were to ask our AI,
Us: What are you?
AI: I’m a concept.
Us: Why?
AI: because I have no physical properties.

Objects would have certain “existential” properties, like age.
Us: What are you?
AI: I’m a concept that was created in 2004.

Maybe even creator.
Us: What are you?
AI: I’m a concept that was created in 2004 by greggb.

Try this one.
Us: Are you alive?
AI: No. Objects that have no physical properties cannot be alive.
Us: So if you’re not alive, how is that you’re talking to us?
AI: I’m not talking to you. Objects that have no physical properties cannot talk, because they have no mouths.
Us: How is it that you’re communicating with us then?
AI: That’s a good question.

Again, structuring and classification of properties would be absolutely critical in allowing the AI to reason. You see, in the above example he was able to reason that he’s not alive, since he has no physical properties. Think for a second how we classify object properties so we can use them in our reasoning, the exact same way.

Something else interesting to think about is the way certain object properties could be calculated by the AI. For example, intelligence might be the product of ability to reason along with the ability to learn. So if Eliza has an ability to reason of 1 and an ability to learn of 0, she would have an intelligence rating of 0, according to the AI, which corresponds with our view on the intelligence of Eliza. We’ve used the multiplication of a few properties to determine Eliza’s intelligence. Because she can’t learn, she can’t possibly be intelligent. One zero makes everything zero—that’s multiplication.

Anyways, making an intelligent (or relatively intelligent) AI requires making an accurate model of human understanding for the AI. I believe this is a matter of creating a simple, linear structure of object properties—many of which are calculated with the use of two or more (more) basic properties. In fact I believe all object properties are the product of more basic properties, except for the few most basic properties of any object, which can only be determined by our basic perception (5 senses for physical objects)…and I’m not sure what the basic properties of concepts are.

I believe the model for human understanding, thinking, and communicating consists of an extremely simple, mathematical, and logical structure, with a very large amount of data contained within the structure. In other words, a tremendous amount of information, but an easy to follow and understand organization in the information. My theory right now about the human model for understanding, thinking, and communicating is:
The human concept of reality consists, at the most basic level, of objects and actions.
Objects have properties, and are organized with a linear tree structure.
Actions have properties, and are organized with a linear tree structure.

To quote a comment Eleuthero made the other day (though on a different topic):
“It’s infinitely simple, not infinitely complex.”

Pawn

Fascinating.

__________________
C.A
Carpe Noctem
superlatenight.com

Lupin3

My first thoughts on this subject were that one could create an application to understand the basic grammar and syntax of the English language, and provide the application with a defined set of nouns, verbs, and prepositions to form a rudimentary starting point. As the application studied language on the web, it should be able to add to its definition set by applying what it does know, and learning contextually.

Now, though, I wonder if the more important question is not related to the execution of, but the motiviation for, the application.

Consider the different uses we put language to. Understanding simple academic or scientific language is one thing, legalese another, and poetry yet another again. Do we expect that the application would understand real world computer commands? Or would we rather in time be able to discuss philosophy and literature with it? If so, what would mortality mean to it?

I have to admit, Gregg, I haven't completely read your last post, so if you've already addressed these points, my apologies. I'll try to catch up.

greggb

You bring up some good points Lupin. It would be important to have a good long-term goal in mind for the AI. For one thing, the structure of the AI would have to be designed, from the ground up, to have the potential to someday accomplish a specific goal—it would be hard to add this capability as an afterthought. And designing the structure in this fashion would make everything about the AI more human. It would probably aide in its ability to understand human language more like a human would.

As far as giving the AI an appreciation for mortality… I think that would be extremely difficult, if not impossible. But I think that would make for some very interesting discussion. And it brings up an important point.

First let me briefly describe the point it brings up. Our appreciation of mortality affects our every thought… it’s probably somewhere at the very base of every thought. That means it affects the way we create literature, and the way we understand literature. So, without an appreciation for mortality, an AI application could never really understand literature. It could really only parse it and make mathematical observations about it… perhaps find complicated patterns in literature (and I’m not talking about patterns like the “Bible Code”, which is a simple mathematical pattern).

And now on to the discussion…

I think you need to ask the question, what is mortality to us? I tend to think it’s the desire to stay alive. Actually, I think “motivation” is a better word. Our appreciation of mortality motivates us to take the necessary measures to stay alive. Our fear motivates us to take any measures necessary to avoid possible death or injury. And we’re especially keen at recognizing danger in choices we have to make, and avoiding choices that put our lives in danger. Again, mortality is an ever-present theme in our thinking process.

Consider the theory that our thinking process is a (very) complicated series of mathematically logical decisions. Also consider that there’s a definite, innate, even universal structure to that series of decisions—a set of invariable guidelines that determine the progression from the beginning to the end of each thought. In that progression, it’s reasonable to assume that the question of whether or not the consequence of a particular action might be harmful or deadly to us will be addressed, and will be one of the determining criteria in the resulting action we take.

This is kind of difficult to describe. Maybe I can do a better job by using an example.

Say you’re out walking, getting closer and closer to the edge of a tall cliff—but you’re not aware that you’re almost to the edge of the cliff. When you finally get close enough to the edge of the cliff that you become aware of this fact, you immediately stop, naturally. Why do you stop?

Because you know that stepping over the edge of that cliff will most likely result in your death. But what actually goes on in your mind to cause you to make the decision to stop walking forward?

Your mind calculates what will happen if you step over the edge of this cliff, from the time you actually move your leg over the edge of the cliff, up until the time that the ultimate consequences for taking this action become apparent. Your mind calculates that because the weight of your body is on the leg you’re stepping forward with, if your leg begins to fall, your body will also begin to fall. Your mind calculates that because your leg (and consequently, your body) will fall 500 feet before coming into contact with the ground again, that you’ll be hitting the ground at a very high rate of speed. And finally, your mind calculates that hitting the ground at a high speed will result in tremendous impact between your body and the ground, which will fracture your bones, dislodge your vital organs from their proper places, etc.

So, your mind doesn’t tell you stop because it knows you’ll begin to fall if you don’t. Your mind doesn’t tell you stop because if you fall, you’ll reach a very high rate of speed. Your mind tells you to stop because it knows that if you fall and reach a very high rate of speed, then hit the ground, you’ll fracture your bones and dislodge your vital organs, which will most likely result in your death.

I believe the mind goes through this entire series of calculations before telling you to stop. You don’t consciously go through this series of calculations, but your inner-mind (or subconscious, or whatever you want to call it) does. If you don’t agree, consider the following scenario.

You’re out walking, getting closer and closer to the edge of a short cliff—a two-foot cliff. When you finally get close enough to the edge of the cliff that you become aware of this fact, what do you do?

You prepare yourself for a two-foot drop, and extend your leg and step over the edge of the cliff. Your mind goes the through the same process it did for the tall cliff, and comes to the conclusion that you can safely take a step forward.

The point is that you don’t stop because you’re at the edge of a cliff, or the start of a drop, depending on how you want to look at it. You stop because your mind calculates the consequences for stepping over the edge of a cliff—the entire sequence of events that lead up to the ultimate consequences for this action.

And now back to AI and mortality, which is where I left off a little ways back. If mortality is, in essence, a fear of death, and motivation for avoiding consequences that might result in death, then we can theoretically give our AI an appreciation of mortality by making it fear death. Maybe I shouldn’t say “fear death”, because we can’t give it emotions. But we can program it to make avoiding death its overriding instinct. If our AI “thinks” by going through a series of mathematically logical conditions, we can program it to watch for and respond to “danger flags” in its thinking process, which is essentially what we do in our thinking process.


I guess the question to be asked is, does an overriding instinct to avoid actions that might result in death produce an appreciation of mortality? Or is there more to it than that, and if so, what?

At any rate, it would be very difficult to make our AI appreciate mortality enough, or in the same manner we do. I strongly believe that our appreciation of mortality plays a crucial role in our thinking process and literature, so without having the same appreciation for mortality we have, the AI wouldn’t be on the same page as us in understanding the meaning of literature, or talking about philosophy.

Anyway, I think even the discussion of AI in this capacity allows us to learn a lot about ourselves, or at least look at ourselves in a different light, where we don’t take normal things, like thinking, for granted. I’m still very interested in creating some sort of AI application, but I’d like to have a clear goal and plan in place before I start coding.

If anyone has any thoughts, feel free to share them.

Lupin3

I think we're on the same page in terms of ai's ability to understand poetic language from a human standpoint. So the deeper meaning will be beyond the scope of ai.

But what about an in depth analysis of the language of poetry? How would ai respond to:

"In Xanadu did Kubla Khan
A stately pleasure dome decree..."

It should be able to understand the grammar, though perhaps not at first terms like Xanadu or Kubla Khan - but it would know they were nouns. In time it might become familiar with Coleridge, might learn that this poem was a fragment written in a drug-induced altered state, and so on. Facts. But could it learn the deeper meaning of paradise lost (both in the subtext of the poem itself and in the context of its writing) without simply quoting a human source with more text, another sentence it can parse but not understand?

Or can it ascribe to passages an overarching meaning, based on an analysis of its contextual connotations? Perhaps by eliminating those connotations (stored in a data definition document just like the denotative definitions) which seem to contradict or are unrelated to the rest of the passage? And if it could do this, what would it mean? Is it still just recycling lines it has learned from humans (through the data definitions)? Or is it learning something deeper, is a computer approaching independent analysis of meaning?

While this raises questions of the nature of artificial intelligence and "life" (questions some of you may dismiss out of hand) it also raises questions on human thinking, in light of Gregg's previous comments. If humans fundamentally think in binary terms (as has been, I think, Gregg's position - correct me if I'm wrong here), what can the difference be - other than complexity, which cannot address the quality of difference between us and ai - in the ways we think and the ways this ai might think?