I borrowed this from a reddit post.  Very pointed questions, but very pertinent.

A.I. is the most failed scientific enterprise in history. Even alchemy at least eventually gave rise to modern chemistry. What exactly is wrong with machine translation and NLP? Is there something about human language we don’t completely understand? Do the researchers of NLP refuse to acknowledge that sentences have a meaning outside of the words?

Thoughts?  Especially regarding mishandling of sentences in user input.

A.I. is the most failed scientific enterprise in history.

That’s debatable, even if you concede failure, which few would do. Successful applications of A.I. are so all-pervasive now that they are practically part of the furniture. On the other hand, fusion is still looking pretty lame unless all you want to do is blow stuff up. If you take that line further then medicine, economics and political science are all failures too because people still die, go broke, and start wars.

Even alchemy at least eventually gave rise to modern chemistry.

How long did that take? Alternatively, how long did it take to achieve sustained controlled flight, and why?

What exactly is wrong with machine translation and NLP?

Good question. Pay me one million euros and I’ll tell you.

Is there something about human language we don’t completely understand?

Yes.

Do the researchers of NLP refuse to acknowledge that sentences have a meaning outside of the words?

No.

The machine translation I cited in this post was pretty good I thought, except for the couple glaring mistakes I noted. But it’s quite usable. (In that post I also noted how I thought it could be improved: have it learn corrections that I can teach it while interacting with the program…)

The quotation reminds me of Hubert Dreyfus (http://en.wikipedia.org/wiki/What_Computers_Can’t_Do). But computers today play chess better than grand champions, recognize faces, drive cars, beat Jeopardy champions…

Sentences have meaning outside the words, but it’s not the most important one :)

As some of you may know, I grew up in Zurich, because my mother studied at the CG Jung Institute, and my father worked with the IBM Zurich Research Lab.  Anyway, long story short, I am now preparing a blog post on “Computational Dreaming”.  The foundation of Computational Dreaming is “Metaphor Analytics”. 

As you may or may not be aware, the United States Intelligence Advanced Research Projects Activity (IARPA) has initiated a “Metaphor Program” in order to crack the non-literal meaning of language.  I believe, the logical conclusion of this work would be no less than to mechanistically tap the collective unconscious.

So, the question is, how will robots dream?

As long as we focus on words, singularity will never be reached.

Real AI will come from an understanding of the world, talking about it should be done after you understand the world. Just like kids expose their intelligence.

Perhaps that’s true, Erwin, but in the case of most chatbots, or other similar technologies, words are pretty much all you have to go by, and so we concentrate on getting the most we can out of understanding them. When w’ve got that “hurdle” cleared, I’m sure that at least some of us will move on to even more complexity, such as integrating visual understanding (yes, I know that there are already some folks doing this, but very few in this community are). and so on.

So what messes up bots the most?  Big, complicated sentences where the subject is ambiguously referred back to?  Grouping words together as a subject when it’s a single subject?  Just curious—I’m more of an AI thinker than a bot programmer.

Erwin Van Lun - Dec 19, 2011:

As long as we focus on words, singularity will never be reached.

Real AI will come from an understanding of the world, talking about it should be done after you understand the world. Just like kids expose their intelligence.

And to do that you must understand understanding.  Digitial computers will never understand the way we do.  I don’t think a digital machine can know the words ‘love’ and ‘hate’ for example.

Computers will have a kind of ‘relational’ understanding of words (by if-then rule chaining, cause & effect, etc).

I would argue that a chat bot ‘understanding’ of words is analogous to a calculator and numbers and operators.  In fact, you could argue that a calculator doesn’t know the concept of ‘5’ as a human mind does, and it doesn’t even add, subtract or multiple or divide the way a human does, the calculator doesn’t even have a clue what dividing means, or *that it is even doing it*. 

But, the result is productive, and a human performing 10/5=2 gets the same results as a computer, the are functionaly equivellent.

The bottom line is, all that matters is functionality.  How it does it, what it is made of, how it works, how it “understands” what numbers, operators, or words in a dictionary are, it doesn’t matter.  Whether that be representing a ‘5’ as 8 bits 00000101.  Or whether performing 2x8 is done by ‘lookup’ the way a human does it, or by turning off and on transistors in silicon in a specific pattern to get the result.

@Dave: I’m suggesting that the chatbot/NLP community is working on words and we’ve been some serious steps forwards lately, but I still believe we are working on components of a car, without having the frame.

The frame is component driven (not containing any language at all).

Witthout having any interaction with other humans, what can we observe using two sensors: two eyes, two ears?
Well, the first
We see colors
We see intensity
We hear tones, high tones, low tones
We hear volume differences

Based on these basic observations, the very first observations of a newborn baby, we are starting to cluster
-which areas of colors belong to one object?
-which object makes which sounds?

The next step is: have we seen this before?
-in another context, on an distance.

If we have not seen this before, does is resemble something we have seen before?
A dog, a cat, a table, a car, a knife, a fork, a ... etcetera.

All this clustering still happens without using any language,

Then, we’ll have observations on processes: how is the world are us changing: A bee flies, A bird flies to. A cat jumps. Cat eats birds.

All simple observations. Very common sense knowledge. And yet still no language.

As soon as we start to talk about observations, then language gets involved.

In today’s AI research however, language is key. And I belief that’s not were the breakthrough will come from.

@Victor: Even humans experience love (or any other emotion) in a total different way. But still, millions of songs and poems are written about this vague concept.

I belief that computers will be able to experience love, as soon as they get spiritual sensors that measure energy waves around humans and models will be available for emotions.

“Spiritual Sensors”? That’s an interesting concept. I think I understand the intent here, but that’s possibly a bit of a stretch I think, even with the present technology. We’ll just have to see.

Sometimes I think that maybe true AI is like the Ecological Emergent Properties.
If the subject is Ecology it’s impossible to study the parts to understand the whole, because when you put all the parts together occurs the Emergent properties and behaviors that can’t be found in the parts apart.
So, maybe joining all the parts of AI and robotic, image recognition, sintax, logics,  artificial feelings ( I think I’ve read something like “lovotic” - or something like this), a sense of feeling it’s suroundings ( maybe by skin sensors, cameras, microphones),  MIT Social cognition and social learn - Google “Robot Leonardo MIT”.
Maybe all this factors together could bring emergent properties to AI…
Just a piece for thought.

I wonder if the evolution of AI will follow a similar path of it’s biological counterpart? Meaning, once logical evolution reaches the point of “self awareness”, then the next phase would be that of a quest for the spiritual meaning of its own existence. Is this possible? I believe it could be when the current digital binary technology moves forward to that of biocircuits that are capable of the same thought processes of humans.

We are really just in the beginning stages of developing artificial intelligence.

Laura Patterson - Dec 29, 2011:

I wonder if the evolution of AI will follow a similar path of it’s biological counterpart?

I think it will.  As soon as it comprehends some aspects of the universe and notices that there is no reason for the universe to even exist it might start thinking about God.

Toby Graves - Dec 29, 2011:

Laura Patterson - Dec 29, 2011:

I wonder if the evolution of AI will follow a similar path of it’s biological counterpart?

I think it will.  As soon as it comprehends some aspects of the universe and notices that there is no reason for the universe to even exist it might start thinking about God.

Quite possibly the answer to why or how the universe became to be will then be answered.

I imagine what kind of role would we play for AI creatures…
“Elder brothers”, “fathers”, “creators”...
If they are supposed to ask about God…

@Erwin ,
maybe not a kind os spiritual sensor ( like a photo kirlian), but maybe future robots can sense our temperature, heart beating, voice level, iris contraction and make an estimative of our “inner state of spirit”. Indeed almost all these sensors already exist, it’s just necessary to put all these devices to work together to achieve this single purpose.
Also, todays AI is already able to make some inferences about human feelings expressed by facial expressions. Take a look at MIT researches ( http://robotic.media.mit.eduprojects/robots/leonardo/overview/overview.html)

Erwin Van Lun - Dec 19, 2011:

As long as we focus on words, singularity will never be reached.

Real AI will come from an understanding of the world, talking about it should be done after you understand the world. Just like kids expose their intelligence.

On the other hand, “In the beginning was the Word…God said let there be light…” So words came first :)