Final Essay

Suppose that you were going to create an AI system to do the following: Given the names of two people, determine from their Wikipedia articles whether they ever met in person.

Your assignment is to write a 2000 word or longer essay sketching what would be involved: how would you design it, what features would it have to have, what kinds of issues would arise, what kinds of problems would be easy, what kinds of problems would be hard, how it could be evaluated, and so on.

The design is largely up to you. In particular, it is your decision whether:

Limited version: If you want, you may restrict the category of people in some way. E.g. two politicians, or two athletes, or two authors. If you plan to do this, please check with me well in advance. Obviously, I will expect a more detailed analysis for this category than for the general problem.

In preparing this assignment, you should look at a number of examples of pairs of individuals, and consider what would be needed to answer the questions for them. For example:

George Washington -- Martha Washington:     Yes. 
George Washington -- Abraham Lincoln:       No 
George Washington -- John Adams:            Yes 
George Washington -- Edward Braddock:       Yes 
George Washington -- Gilbert Stuart:        Yes 
George Washington -- Wolfgang A. Mozart:    Probably not 
     (actually, certainly not, but I don't think that can be determined from
     the Wikipedia articles)
George Washington -- Benjamin Franklin:     Yes, though from the Wikipedia 
     articles it would hard to say more than that it is very probable. 
Giuseppe Verdi -- Richard Wagner:           No
Ludwig van Beethoven -- Wolfgang A. Mozart: Yes
Emperor Komei [of Japan] -- Commodore Matthew Perry:  No.
Emily Dickinson -- Laura Ingalls Wilder:    Almost certainly not.
Oscar Wilde -- Walt Whitman:                Yes.
John Updike -- Saul Bellow:                 Almost certainly.
John Updike -- Francis Crick:               Probably not.
Francis Crick -- John Steinbeck:            Yes
     (at the 1962 Nobel prize ceremony)
John Havlicek -- Kareem Abdul-Jabbar:       Yes
Rod Laver -- Ken Rosewall:                  Yes
Rod Laver -- Roger Federer:                 Almost certainly.
Arthur Ashe -- Roger Federer:               Almost certainly not.

Note: Some of these are extremely easy; for example, the article on Oscar Wilde says, ``Wilde also met Walt Whitman ...'' George Washington died before Abraham Lincoln was born, so that can easily be answered "No". (If queries are picked at random from the space of all people, then a very large fraction of queries can be answered "No" just on that basis.) Others would require an very sophisticated (by the standards of AI programs) understanding of social realities. For instance, inferring from the Wikipedia articles that John Updike probably met Saul Bellow requires knowing in the twentieth century, any two leading figures in the same line of work (novelist) from the same country have probably met. Your paper should certainly describe how the proposed program could solve easy cases. You should discuss some more difficult cases, but which difficult cases you want to deal with is up to you.

Your paper can analyze any or all of the above examples, or you can find your own examples, but you should certainly analyze some number of specific examples.

What I expect

Obviously, this problem is extremely open-ended, and it would not be possible, in the current state of the art, to develop a system that could actually handle all examples of difficulty comparable to some of those I've listed above. Therefore I'm not expecting a description of a system that will solve the problem in all cases. What I want is a fairly realistic description of how you would build a reasonably good system of this kind if, say, you were hired by Google or Microsoft to do it, with a budget that enabled you to hire a team of a dozen crack AI programmers over a five year period.

Your system may, and should, use techniques and algorithms that we have discussed in class. However, you should not spell out the details of these, unless you want to propose an original technique/algorithm which for some reason is particularly suited to this particular problem. Similarly, if your system requires solving an AI problem that is known and not wholly solved, then you should mention the problem, but you should not go into detail about how it is solved unless you have something new to say about it in this context. For instance, doing this well will obviously require anaphoric resolution (e.g. the article on Verdi says of Verdi and Wagner "They never met".) You should mention this, but you should not regurgitate the standard techniques for anaphoric resolution, unless you think that there is something distinctive about doing anaphoric resolution for this application, or for Wikipedia pages.

However, it is very important that you should give specific examples of rules that you would use in a knowledge-based system or of attributes in a learning system, and demonstrate how they would apply to specific examples.

A discussion of the issue of evaluation (how you would measure the correctness of the program) is welcome but not required.

Deliverable

Either email (to me, not the grader) or hand in printed hard copy of your paper by the due date, December 11 at class time. Handwritten copy will not be accepted. Late submissions will be penalized 10 points out of 100.

If you want you can email me a first draft of your paper, not later than November 27, and I will send back suggestions for improvements. However, I must receive the final version by the due date.

Research and citations

There is no need to do any research for this paper, in the sense of reading additional technical material, and it is unlikely to be very helpful, though it is certainly allowed. Ideas and techniques that is in the textbook or that was discussed in class can all be considered ``well-known'' within the AI community, and therefore need not be given citations. (Actually, there is some technical material in the textbook which, if I were using in a published article, I would give a citation for; but you need not worry about that.) Therefore, it will probably unnecessary for you to have any citations in your paper, except for the Wikipedia articles that you quote for examples. However:

Getting Help

I shall be happy to discuss this assignment with you at almost any length you want. So if you are at all confused about any part, or all, of the assignment, or you want to bounce ideas off of me, by all means send me email, or come to my office hours, or, if you think you have a question of general interest, ask during class time. Do not spend long periods of time confused or lost or uncertain how to proceed without consulting with me.

Final Warning

What I want to see in this paper is the fruit of serious, substantive thought. Therefore, you should start thinking about and working on this project well in advance of the due date. Do not start this on the night of December 10, if you want to do well.