Tuesday, September 26, 2006

The Saxena Test

Has my computer turned into a human?

This question is troubling humanity for quite some time now. Many people have tried to answer it by giving our beloved computer tasks that only a human can be expected to complete, most of which have historically been intelligent tasks. One of the most famous such test is the Turing Test. According to Alan M. Turing, a computer can be considered intelligent (and consequently human) only if it passes the Turing Test. Here's the deal:

You are sitting on one side of a curtain. The proposed human being (PHB) is on the other side. You can interact with it through a neutral common language, say printed questions and answers. If it is not possible to determine if the PIB on the other side is a human or a computer, then a computer that is able to remain undetected by this test would be considered intelligent. The test is very simple, but till date no computer has been able to pass this test. Numerous attempts have been made to develop intelligent computers, based on humongous memory background. But since these have not been extensive, and thus, these machines have failed to pass the test. So is this test the holy grail, and should our search end here?

Not quite. The rate of advancement in computer sciences is extremely rapid, and predictions that such a computer is just round the corner are rife. This brings us to the fundamental question: What task (or test) can be performed by a human being, and not a computer? Furthermore, if we can frame a test that also distinguishes a human by evaluating human traits other than intelligence, it would definitely be more desirable. Such a test should have the capability to differentiating how a human thinks, as against the database searching capability of a computer. Most of the questions asked till now have been based on soft computing theories. But with even soft-computing approach applied to computer's thinking, these methods have limited application. Based on the discussions above, it is clear that to frame the ultimate question, we should concentrate on topics where the computer can't be applied rather than where computer is traditionally "slower" than humans. One such task that I thought about is the ability to understand Jokes! I call it "The Saxena Test".

The idea is simple. To the PHB, present a plain text of English prose and ask whether it considers the prose a valid humorous prose (i.e. a joke), and preferably rate it on a scale 1~10 how much humorous does the PIB find it. The beauty of this question is that computers can't be taught what jokes are, it is something we have in our instincts. Although we differ on how much humorous we find certain jokes, there is an obvious pattern of rating among rating too humorous jokes, and less humorous ones. Also, it is possible to tell if there is any joke in the text or not.

Since the set of jokes in an infinite set, when a computer would be developed to try and solve this problem, it is clear that instead of building an exhaustive database of jokes, it would be taught recognizing the language patterns in the joke. But here also, since the exceptions are so many, a computer would definitely fail. As an example, consider this prose:

"Man 1: Hey, would you come to play golf with me?
Man 2: Yes, but only if you call me once.
"

It is clear (at least to me) that the sentence above is not a joke. But it is easy to recollect innumerable jokes that fit into this pattern.

On many occasions, the humor in the joke is dependent on extra information not present in the joke. For example, consider this example:

Two atoms were coming around a corner, where they collide against each other and fall down.
The first atom asks the other: Are you all right?
The second atom replies: I think I have lost an electron.
First atom: Are you sure?
Second atom: Yes, I am positive.

To those who don't understand atomic structure, this wouldn't appear like a joke. Let us take another example:

Descartes is sitting in a bar, having a drink. The bartender asks him if he would like another. "I think not," he says, and vanishes.

It is possible for us to use these cases to our advantage while testing the PHB. Thus after giving one of the standard task of understanding an English prose, a book on history of logic, the computer should start considering the prose above as a joke, and the joke-rating of the prose above should rise substantially.

The next task is quantification of whether the PHB is human or not. The joke-rating provided by the PHB should be normalized and compared with ratings given by actual human beings. If there is high positive correlation, say around 0.8~0.9, then the PHB can be considered as a human.

I thought about this test years back, and thought it was high time I publish it; before someone else does! Neat?