As a digital computer competes to become the most successful player of Jeopardy, this might be an auspicious time to consider the relationship between mind and machine, and the issues raised by IBM's Watson.
Consider the skills that Watson needs in order to succeed. Take a typical Jeopardy question, "287 miles long, it flows past Washington & induces fever in ambitious politicians". Answer, "What is the Potomac River?". First, Watson must process the sentence in order to understand the question. Second, Watson must search a knowledge base and retrieve relevant facts. Third, it must use those facts to generate possible answers. Fourth, Watson must choose the best answer, and determine if it has sufficient confidence in that answer to press the button. Finally, it must generate its answer in the form of a question.
Each of these tasks presents separate challenges. However, we must recall that the designers of Watson were concerned exclusively with maximizing its performance, not necessarily with simulating human performance. So Watson might do very well at jeopardy, but tell us little that is of interest to a cognitive scientist (beyond the fact that a machine can do this sort of thing).
The last task, turning the answer into a question, is undoubtedly the easiest. Although a human might get distracted by focussing on the first four tasks and forget the final twist, this is something a machine would never do. Once one has identified the key fact ("This is the Potomac River"), all that is needed is to apply a rule for transforming statements into questions. This is mostly a rule for selecting the substitute for "This" ("Who?", "What?","When?", etc.)
Judging from previous work in artificial intelligence, the most difficult task is probably the first - the understanding component. It is also the task where, in order to succeed, Watson must probably be most like a human. Even here, though, the task is more straight forward than it is in many natural language contexts. Probably a key starting point is to find a pronoun - "it" in this case - and extract a simple proposition, "it flows". Alternatively, there may be a possessive, "his" or "hers" ("After his circumnavigation of the world he was made mayor of Plymouth, England"), or "this" ("Easy-Off Oven & Grill Cleaner comes in this type of can").
From "it flows", we now know that the original question concerns something that can flow, thus restricting the relevant knowledge to rivers, words, maple syrup, etc. More propositions serve to further narrow the scope: "It is 287 miles long", "It induces fever in politicians", etc. Of course, extracting these propositions is no mean feat, so Watson must possess some significant language skills. And there are always a few questions that require a different approach - "A fur wrap, or what you did if you illegally took one", for example. In this last example, there is the tricky problem of determining what "one" refers to.
Watson can now move on to the search step. A straight forward associational process will suffice: "flow" » "river", "river" + "Washington" » "Potomac", "politician" + "fever" » "Potomac fever". Later in the course we'll see how this can be done. The real challenge is to design the knowledge base in a way that permits rapid search. Here is where Watson's enormous memory and speed give it an advantage.
Note that in order to succeed, Watson does not have to "understand" much of what it has generated from the question, or of its knowledge. Consider an oft-quoted error of Watson's: "What do grasshoppers eat?" "Kosher". Watson found an association between grasshoppers and kosher food (grasshoppers are indeed a kosher food). A human would never make such a mistake, because we also know that only people eat kosher food. Watson probably also knows this, but does not realize that in this context it is a very important item of information. It should remind you of the "frame"problem discussed earlier in the course.
A tricky problem arises when the search process leads to more than one possible answer, or to an answer that seems to be less than perfect. The task here is to find a decision rule that will maximize correct answers while minimizing false alarms. This is a well known problem in statistical decision theory, and it finds application in many areas of human behavior.
Here are links to some recent articles that discuss Watson's abilities and its place in the development of artificial intelligence:
A discussion of Watson on NPR and this radio broadcast.
A summary of relevant opinions in Tech News Daily
An interview with author Stephen Baker, who has written a book about Watson.
My own conclusion: Watson represents an impressive product of expert knowledge in several different areas that are relevant to human performance. It is a significant achievement in artificial intelligence. Whether or not it adds to our understanding of human behavior remains to be seen. Although it may win at Jeopardy, it would probably not do well on the Turing test.